Flashinglf/awesome-multi-agent-communication-collaboration-evolution

Python

1 commits

updated Jun 24, 2026

README

Awesome Multi-Agent Communication, Collaboration, and Evolution

A curated and updateable repository for research on multi-agent communication, collaboration, robustness, and evolution, focused on LLM-based multi-agent systems and long-tail task assistants.

Papers: 108
GitHub projects and systems: 79
Technical docs, blogs, and field reports: 47
Non-GitHub project/reference rows: 4
Core papers: 42
Core GitHub projects/systems: 17
Core technical readings: 7
Core non-GitHub references: 1
Papers with non-arXiv primary source: 41
Papers with verified venue metadata: 53
Last verified: 2026-06-13
Language: English | 中文

Featured Technical Readings

Anthropic - How we built our multi-agent research system: Anthropic 对生产级 research multi-agent system 的工程复盘，包含 lead agent、并行 subagents、任务分解、trace 和 citation 验证经验。
Anthropic - Building multi-agent systems: 讨论什么时候值得使用多 agent，以及 context protection、parallelization、specialization 三类收益和高额 token 成本之间的取舍。
Anthropic - Multi-agent coordination patterns: 总结 generator-verifier、orchestrator-subagent、agent teams、message bus、shared state 五类常见协作模式及适用条件。
Anthropic - Building Effective AI Agents: Anthropic 对 workflows 与 autonomous agents 的实用区分，并总结 prompt chaining、routing、parallelization、orchestrator-workers、evaluator-optimizer 等模式。
OpenAI Agents SDK - Orchestrating Multiple Agents: OpenAI Agents SDK 官方多 agent 编排文档，覆盖 agents-as-tools、handoffs 和 manager 模式。
OpenAI Cookbook - Orchestrating Agents: OpenAI Cookbook 示例，展示如何用 routines、handoffs 和工具组织多个 agents。
OpenAI - Harness engineering: OpenAI 对 harness engineering 的实践总结，强调通过约束、验证和工具环境让 agent-first 软件更可靠。
OpenAI - Building more helpful agents with a new evaluation framework: OpenAI 发布 BrowseComp，用难以搜索定位的问题评估 agent 深度检索和信息整合能力。
LangChain - How and when to build multi-agent systems: LangChain 对 multi-agent supervisor、handoff、subagent 工具化等模式的实践建议。
AutoGen Documentation - Multi-agent Design Patterns: AutoGen 官方文档中的多 agent 设计模式入口，覆盖 Core runtime 下的协作组织方式。

Scope
Featured Technical Readings
Main Documents
Category Overview
Source Coverage
Core Papers
Core GitHub Projects and Systems
Core Technical Readings
Core Non-GitHub References
Complete Paper Catalog
GitHub Project Catalog
Technical Docs, Blogs, and Field Reports
Non-GitHub Project and Benchmark References
Maintenance

Scope

This repository is intentionally narrower than generic agent-framework lists. It tracks work that helps answer three questions:

Communication: who should receive what information, when, and under what token budget?
Collaboration: how should agents form teams, coordinate, verify, and recover from failures?
Evolution: how can traces and feedback improve prompts, roles, tools, edges, topologies, and workflows?

Main Documents

Category Overview

Catalog	Category	Entries
Papers	Communication & Routing	22
Papers	Task-Oriented Collaboration & Dynamic Teaming	22
Papers	Debate, Consensus & Verification	16
Papers	Role-Playing, Agent Society & Simulation	4
Papers	Software Engineering Multi-Agent Collaboration	9
Papers	Robustness & Reliability	7
Papers	Evolution & Optimization	16
Papers	Evaluation & Benchmarks	8
Papers	Surveys & Taxonomies	4
GitHub projects	Frameworks & Runtimes	26
GitHub projects	Protocols & Interoperability	7
GitHub projects	Benchmarks & Evaluation	16
GitHub projects	Observability & Operations	7
GitHub projects	Research Implementations & Baselines	18
GitHub projects	Reference Systems & Case Studies	5
Technical docs/blogs	Anthropic Official Engineering Articles	14
Technical docs/blogs	OpenAI Official Engineering Articles and Docs	8
Technical docs/blogs	Multi-Agent Framework Documentation	7
Technical docs/blogs	Protocol Documentation	3
Technical docs/blogs	Evaluation, Runtime, and Engineering Blogs	11
Technical docs/blogs	High-Quality Personal and Community Blogs	4
Non-GitHub references	Protocols & Interoperability	1
Non-GitHub references	Benchmarks & Evaluation	1
Non-GitHub references	Observability & Operations	1
Non-GitHub references	Research Implementations & Baselines	1

Source Coverage

Primary source	Papers
arxiv	67
openreview	16
acl	10
neurips	4
pmlr	4
aaai	2
journal	2
acm	1
ieee	1
ijcai	1

Core Papers

Area	Paper	Year	Venue / source	Why it matters
Communication & Routing	RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory	2025	arxiv	直接命中项目“通信代价大、关键信息传不准”的问题，可作为预算约束上下文路由主 baseline。
Communication & Routing	Improving Multi-Agent Debate with Sparse Communication Topology	2024	arxiv	支撑“不是所有 agent 都应该互相通信”的设计假设，适合做全连接通信的对照实验。
Communication & Routing	Dynamic Generation of Multi-LLM Agents Communication Topologies with Graph Diffusion Models	2025	arxiv	适合支撑“动态 agent 组队 + 动态通信网络”的算法路线。
Communication & Routing	Towards Adaptive, Scalable, and Robust Coordination of LLM Agents: A Dynamic Ad-Hoc Networking Perspective	2026	arxiv	同时覆盖通信路由、动态协作和鲁棒性，是项目基础设施设计的强相关论文。
Task-Oriented Collaboration & Dynamic Teaming	A Dynamic LLM-Powered Agent Network for Task-Oriented Agent Collaboration	2023	arxiv	可作为动态组队和 agent 选择机制的早期 baseline。
Task-Oriented Collaboration & Dynamic Teaming	AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation	2023	arxiv	是 LLM 多 agent conversation framework 的基础论文，应作为协作运行时和对话协议 baseline。
Task-Oriented Collaboration & Dynamic Teaming	AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors in Agents	2023	ICLR 2024 Poster	补齐早期通用 LLM 多 agent 协作框架，对动态组队和正负涌现行为分析有参考价值。
Task-Oriented Collaboration & Dynamic Teaming	Talk Structurally, Act Hierarchically: A Collaborative Framework for LLM Multi-Agent Systems	2025	arxiv	直接补强“通信协议 + 分层协作”的设计路线，可作为动态团队协作的强 baseline。
Task-Oriented Collaboration & Dynamic Teaming	Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks	2024	arxiv	是 orchestrator-specialist 架构的强参考系统，也提供 AutoGenBench 这类可隔离评测工具。
Robustness & Reliability	On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents	2024	ICML 2025 Poster	直接对应项目“agent 上下线、能力波动、整体任务执行不稳定”的痛点。
Robustness & Reliability	VerifyMAS: Hypothesis Verification for Failure Attribution in LLM Multi-Agent Systems	2026	arxiv	自进化系统必须先知道失败来自分解、路由、执行、合成还是验证，故强相关。
Robustness & Reliability	Maris: A Formally Verifiable Privacy Policy Enforcement Paradigm for Multi-Agent Collaboration Systems	2025	arxiv	直接补强多 agent 通信的治理层，适合支撑敏感信息跨 agent 泄露和权限边界设计。
Task-Oriented Collaboration & Dynamic Teaming	Towards Effective GenAI Multi-Agent Collaboration: Design and Evaluation for Enterprise Applications	2024	arxiv	与项目落地到任务助手和企业级复杂任务高度贴合，尤其适合作端到端可用率与成本指标参考。
Task-Oriented Collaboration & Dynamic Teaming	Scaling Large Language Model-based Multi-Agent Collaboration	2024	ICLR 2025 Poster	是协作规模化和拓扑组织的重要论文，直接支撑“多个开源榜单 SOTA”和大规模协作框架目标。
Debate, Consensus & Verification	Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate	2023	EMNLP 2024	是 debate/verification 协作主线的基础论文，已有条目保留为核心协作机制。
Debate, Consensus & Verification	Improving Factuality and Reasoning in Language Models through Multiagent Debate	2023	arxiv	是 LLM multiagent debate 的代表性基础工作，应作为 debate 协作机制和 factuality 改进 baseline。
Debate, Consensus & Verification	ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs	2023	ACL 2024	对异构 agent 的 consensus 机制、置信度字段和投票聚合策略很有参考价值。
Debate, Consensus & Verification	Mixture-of-Agents Enhances Large Language Model Capabilities	2024	ICLR 2025 Spotlight	是多模型/多 agent 层级聚合的关键 baseline，也暴露了后续稀疏路由和成本控制需求。
Debate, Consensus & Verification	GroupDebate: Enhancing the Efficiency of Multi-Agent Debate Using Group Discussion	2024	arxiv	直接对应项目“协作不能增加 token/耗时”的成本约束。
Debate, Consensus & Verification	Stay Focused: Problem Drift in Multi-Agent Debate	2025	EACL 2025	重要负面结果，提醒协作轮次越多不一定越好，必须有终止和聚焦机制。
Debate, Consensus & Verification	iMAD: Intelligent Multi-Agent Debate for Efficient and Accurate LLM Inference	2025	AAAI 2026	非常贴合“按需协作”和“成本不增”的产品目标。
Debate, Consensus & Verification	ARMOR-MAD: Adaptive Routing for Heterogeneous Multi-Agent Debate in Large Language Model Reasoning	2026	arxiv	与项目的“能力波动、动态路由、成本控制”高度相关。
Software Engineering Multi-Agent Collaboration	ChatDev: Communicative Agents for Software Development	2023	ACL 2024	是软件工程多 agent 协作的代表性工作，应作为固定流程和语言通信协作 baseline。
Software Engineering Multi-Agent Collaboration	Co-Saving: Resource Aware Multi-Agent Collaboration for Software Development	2025	arxiv	直接服务项目“成本不增”和“从历史轨迹进化”的目标。
Software Engineering Multi-Agent Collaboration	Optimizing LLM-Based Multi-Agent System with Textual Feedback: A Case Study on Software Development	2025	arxiv	是“根据失败解释优化多 agent 角色/prompt”的直接相关论文。
Software Engineering Multi-Agent Collaboration	Self-Organizing Multi-Agent Systems for Continuous Software Development	2026	arxiv	强相关于动态 agent 组队、长期任务和自组织协作。
Software Engineering Multi-Agent Collaboration	Towards Engineering Multi-Agent LLMs: A Protocol-Driven Approach	2025	APSEC 2025	与“通用协作框架”和 A2A/MCP 风格协议化设计高度贴合。
Evolution & Optimization	Cross-Task Experiential Learning on LLM-based Multi-Agent Collaboration	2025	arxiv	直接补强“跨任务经验积累”和“协作系统进化”方向。
Evolution & Optimization	GPTSwarm: Language Agents as Optimizable Graphs	2024	ICML 2024 Oral	几乎直接对应“多 agent 反馈链路如何优化”，可作为群体进化主 baseline。
Evolution & Optimization	AFlow: Automating Agentic Workflow Generation	2024	ICLR 2025 Oral	适合作离线 workflow 自动优化器，对项目“可自进化协作系统”非常关键。
Evolution & Optimization	EvoAgent: Towards Automatic Multi-Agent Generation via Evolutionary Algorithms	2024	NAACL 2025	与项目“群体进化”命名和目标直接对齐。
Evolution & Optimization	Multi-Agent Collaboration via Evolving Orchestration	2025	NeurIPS 2025	直接补齐“协作编排本身可进化”的顶会工作，适合和 GPTSwarm、AFlow、EvoAgent、EvoMAS 共同作为进化主线 baseline。
Evolution & Optimization	Evolutionary Generation of Multi-Agent Systems	2026	arxiv	对工程落地尤其重要，因为配置空间进化比任意代码进化更可控。
Evolution & Optimization	Evolve as a Team: Collaborative Self-Evolution for LLM-based Multi-Agent Systems	2026	arxiv	最贴合“多 agent 之间互相影响的复杂反馈链路”的研究问题。
Evolution & Optimization	Swarm Skills: A Portable, Self-Evolving Multi-Agent System Specification	2026	arxiv	对“通用协作框架、可迁移到其他智能体产品线”非常贴合。
Evaluation & Benchmarks	Agent-as-a-Judge: Evaluate Agents with Agents	2024	ICML 2025 Poster	自进化需要 reward/critic，该论文可支撑反馈信号设计。
Communication & Routing	Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems	2025	ICLR 2025	直接命中“协作不能显著增加 token/成本”的工程边界，是通信剪枝和经济型协作拓扑的关键顶会论文。
Communication & Routing	G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks	2025	ICML 2025	与动态通信网络、任务自适应路由和低成本协作高度相关，可与 GPTSwarm、AgentPrune、MasRouter 形成拓扑优化主线。
Communication & Routing	MasRouter: Learning to Route LLMs for Multi-Agent Systems	2025	ACL 2025	直接服务“谁该参与、何时参与、以什么成本参与”的路由问题，是多 agent 通信与动态组队的关键 ACL 论文。
Task-Oriented Collaboration & Dynamic Teaming	Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation	2025	AAAI 2026 Oral	直接补强自动组队与通信拓扑生成，是从固定团队走向按任务生成团队的高质量前沿工作。
Evolution & Optimization	MAS-GPT: Training LLMs to Build LLM-based Multi-Agent Systems	2025	ICML 2025 Poster	直接面向“根据任务自动生成协作系统”的进化方向，可与 AFlow、SwarmAgentic、ARG-Designer 共同构成自动化设计主线。
Evaluation & Benchmarks	MultiAgentBench: Evaluating the Collaboration and Competition of LLM agents	2025	ACL 2025	是当前多 agent 协作评测的重要 ACL 基准，可用于验证通信拓扑和协作策略是否真正提升任务完成与协作质量。

Core GitHub Projects and Systems

Area	Project	Why it matters
Frameworks & Runtimes	LangGraph	适合做多 agent 协作运行时和 traceable workflow baseline。
Frameworks & Runtimes	AutoGen	适合参考多 agent conversation、事件驱动通信和分布式 agent runtime。
Frameworks & Runtimes	AgentScope	与“通用协作框架 + 动态协作 + 鲁棒部署”高度贴合。
Frameworks & Runtimes	CAMEL	适合做多 agent 实验平台和 agent society baseline。
Frameworks & Runtimes	ChatDev	是软件工程多 agent 协作和协作 scaling law 的核心参考实现。
Protocols & Interoperability	Model Context Protocol	多智能体框架若要跨产品线复用，工具/上下文接入层建议兼容 MCP。
Protocols & Interoperability	Agent2Agent Protocol	与“通用协作框架”和跨产品线 agent 协作高度相关。
Benchmarks & Evaluation	Multi-Agents-Debate	是 debate 协作机制的核心 baseline。
Research Implementations & Baselines	TalkHier	适合做“结构化消息协议 + 分层协作修正”的核心 baseline。
Research Implementations & Baselines	ReConcile	可作为异构 agent 共识、置信度表达和投票聚合的基础实现。
Research Implementations & Baselines	GPTSwarm	是群体进化和“多 agent 反馈链路优化”的核心实现，应作为主 baseline。
Reference Systems & Case Studies	Magentic-One	是 orchestrator-subagent 架构、错误恢复和隔离评测的高价值参考系统。
Research Implementations & Baselines	AgentPrune	是低成本通信和通信鲁棒性实验的重要 baseline，适合直接纳入复现实验。
Research Implementations & Baselines	MasRouter	是动态路由、低成本协作和模型选择实验的核心 baseline。
Benchmarks & Evaluation	MARBLE / MultiAgentBench	对验证多 agent 通信拓扑、协作协议和竞争/协作行为质量非常关键。
Research Implementations & Baselines	MAS-GPT	是自动化多 agent 系统生成和进化方向的重要实现。
Research Implementations & Baselines	ARG-Designer	对自动组队、自动拓扑设计和低成本协作非常关键，适合作为前沿 baseline。

Core Technical Readings

Group	Reading	Why it matters
Anthropic Official Engineering Articles	Anthropic - How we built our multi-agent research system	是“多步骤信息搜集+整合”场景最直接的公开工程案例，应作为系统设计和成本评估的必读材料。
Anthropic Official Engineering Articles	Anthropic - Building Effective AI Agents	是多 agent 协作模式选择的基础阅读，能防止把所有任务都错误地做成复杂多 agent。
Anthropic Official Engineering Articles	Anthropic - Building multi-agent systems	直接约束本项目“token 消耗和端到端耗时维持不增”的工程边界。
Anthropic Official Engineering Articles	Anthropic - Effective context engineering for AI agents	多 agent 通信路由本质上依赖上下文预算和信息选择，这篇可直接支撑通信层设计。
Anthropic Official Engineering Articles	Anthropic - Multi-agent coordination patterns	可作为本仓库协作分类和系统架构章节的工程模式骨架。
OpenAI Official Engineering Articles and Docs	OpenAI - Harness engineering	为多 agent 工程化提供 harness、验证和可靠性约束，是系统落地的重要方法论。
OpenAI Official Engineering Articles and Docs	OpenAI Agents SDK - Orchestrating Multiple Agents	对用 OpenAI Agents SDK 复现 supervisor/manager/handoff 协作模式非常关键。

Core Non-GitHub References

Area	Reference	Why it matters
Protocols & Interoperability	SEMAP	虽主要是论文方案，但非常适合作为多 agent 协作协议设计参考。

Complete Paper Catalog

Communication & Routing

Paper	Year	Venue / source	Status	Tags	Why it matters
Learning Multiagent Communication with Backpropagation	2016	NeurIPS 2016	foundational	commnet, differentiable-communication, marl	是“通信可以被学习”的基础论文，可作为 LLM 多 agent 路由学习的理论背景。
Learning to Communicate with Deep Multi-Agent Reinforcement Learning	2016	NeurIPS 2016	foundational	rial, dial, communication-learning, marl	支撑“通信协议可通过环境反馈优化”的理论基础。
Learning Attentional Communication for Multi-Agent Cooperation	2018	NeurIPS 2018	important	attentional-communication, marl, message-routing, cooperation	是“通信不是全连接广播，而是按需要形成通信组”的早期顶会基础论文。
Learning when to Communicate at Scale in Multiagent Cooperative and Competitive Tasks	2018	ICLR 2019	foundational	ic3net, communication-gating, scalable-marl	可借鉴为 LLM agent 通信开关、预算门控和 token 节省机制。
TarMAC: Targeted Multi-Agent Communication	2018	ICML 2019	foundational	targeted-communication, attention, marl	与项目中的“高效信息路由机制”高度同构。
Efficient Multi-Agent Communication via Shapley Message Value	2022	IJCAI 2022	important	message-value, shapley-value, communication-efficiency, marl	为“哪些消息值得发”提供可解释价值估计思路，适合连接 token budget 和通信收益评估。
Exchange-of-Thought: Enhancing Large Language Model Capabilities through Cross-Model Communication	2023	arxiv	important	cross-model-communication, relay, debate, memory	适合整理多 agent 通信协议设计空间。
A Scalable Communication Protocol for Networks of Large Language Models	2024	arxiv	supporting	communication-protocol, llm-networks, scalability, message-passing	为 A2A/ANP/ACP 等工程协议之外的学术协议设计提供参考，适合补齐协议层文献。
Improving Multi-Agent Debate with Sparse Communication Topology	2024	arxiv	core	sparse-topology, debate, communication-cost	支撑“不是所有 agent 都应该互相通信”的设计假设，适合做全连接通信的对照实验。
Learning Multi-Agent Communication from Graph Modeling Perspective	2024	ICLR 2024 Poster	important	learnable-graph, communication-topology, marl, graph-modeling	直接支撑通信拓扑可学习这一方向，可与 LLM 动态通信拓扑论文形成传统 MARL 到 LLM-MAS 的脉络。
SMoA: Improving Multi-agent Large Language Models with Sparse Mixture-of-Agents	2024	PAKDD 2025	important	sparse-mixture-of-agents, response-selection, early-stopping, role-diversity	是 MoA 类系统从全连接通信走向稀疏通信的直接参考，适合做成本受限多 agent 推理 baseline。
Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems	2025	ICLR 2025	core	agentprune, communication-pruning, token-efficiency, adversarial-robustness	直接命中“协作不能显著增加 token/成本”的工程边界，是通信剪枝和经济型协作拓扑的关键顶会论文。
Dynamic Generation of Multi-LLM Agents Communication Topologies with Graph Diffusion Models	2025	arxiv	core	topology-generation, graph-diffusion, cost-quality-robustness	适合支撑“动态 agent 组队 + 动态通信网络”的算法路线。
G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks	2025	ICML 2025	core	topology-design, graph-neural-network, task-adaptive-routing, token-efficiency	与动态通信网络、任务自适应路由和低成本协作高度相关，可与 GPTSwarm、AgentPrune、MasRouter 形成拓扑优化主线。
MasRouter: Learning to Route LLMs for Multi-Agent Systems	2025	ACL 2025	core	routing, llm-selection, multi-agent-systems, inference-efficiency	直接服务“谁该参与、何时参与、以什么成本参与”的路由问题，是多 agent 通信与动态组队的关键 ACL 论文。
RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory	2025	arxiv	core	context-routing, role-aware, structured-memory, token-budget	直接命中项目“通信代价大、关键信息传不准”的问题，可作为预算约束上下文路由主 baseline。
Differentiable Mixture-of-Agents Incentivizes Swarm Intelligence of Large Language Models	2026	arxiv	important	differentiable-routing, sparse-activation, test-time-adaptation, swarm-intelligence	连接“动态通信拓扑”和“在线自适应”，适合作为可学习路由和群体智能方向的前沿参考。
RADAR: Redundancy-Aware Diffusion for Multi-Agent Communication Structure Generation	2026	arxiv	important	redundancy-aware, diffusion, communication-structure-generation, topology-generation	补强 G-Designer、ARG-Designer 之后的通信结构生成路线，适合跟踪 2026 年前沿拓扑生成方法。
RouteMoA: Dynamic Routing without Pre-Inference Boosts Efficient Mixture-of-Agents	2026	arxiv	important	mixture-of-agents, dynamic-routing, model-selection, cost-latency	对“协作不能显著增加 token 和耗时”的约束非常关键，可支撑预推理前的 agent/model 路由。
TCAndon-Router: Adaptive Reasoning Router for Multi-Agent Collaboration	2026	arxiv	supporting	adaptive-routing, reasoning-router, collaboration, dynamic-selection	与 MasRouter、RCR-Router、ARMOR-MAD 共同构成动态路由方向，可作为后续候选 baseline。
The Five Ws of Multi-Agent Communication: Who Talks to Whom, When, What, and Why -- A Survey from MARL to Emergent Language and LLMs	2026	TMLR 2026	survey	survey, communication, marl, emergent-language, llm-agents	是本仓库通信章节的高质量分类法补充，可帮助把传统可学习通信和 LLM 自然语言通信放到同一脉络。
Towards Adaptive, Scalable, and Robust Coordination of LLM Agents: A Dynamic Ad-Hoc Networking Perspective	2026	arxiv	core	publish-subscribe, reputation, dynamic-networking, intent-routing	同时覆盖通信路由、动态协作和鲁棒性，是项目基础设施设计的强相关论文。

Task-Oriented Collaboration & Dynamic Teaming

Paper	Year	Venue / source	Status	Tags	Why it matters
Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments	2017	NeurIPS 2017	important	marl, actor-critic, mixed-cooperative-competitive, collaboration	是多智能体协作/竞争学习的重要基础论文，可作为 LLM agent 团队协作前的 MARL 基线背景。
Counterfactual Multi-Agent Policy Gradients	2018	AAAI 2018	important	marl, credit-assignment, counterfactual-baseline, cooperation	为多 agent 协作中的责任归因、失败归因和团队级优化提供基础算法背景。
QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning	2018	ICML 2018	important	marl, value-factorization, team-reward, cooperation	补齐协作任务中 team reward 和 credit assignment 的基础方法，有助于定义多 agent 协作评测指标。
Actor-Attention-Critic for Multi-Agent Reinforcement Learning	2019	ICML 2019	important	marl, attention, actor-critic, collaboration	与 LLM 多 agent 中“只关注相关 agent 输出”的思想一致，可作为协作注意力机制的基础参考。
A Dynamic LLM-Powered Agent Network for Task-Oriented Agent Collaboration	2023	arxiv	core	dynamic-agent-network, agent-selection, collaboration	可作为动态组队和 agent 选择机制的早期 baseline。
AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors in Agents	2023	ICLR 2024 Poster	core	agentverse, dynamic-composition, emergent-behavior, collaboration	补齐早期通用 LLM 多 agent 协作框架，对动态组队和正负涌现行为分析有参考价值。
AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation	2023	arxiv	core	autogen, multi-agent-conversation, tool-use, human-in-the-loop	是 LLM 多 agent conversation framework 的基础论文，应作为协作运行时和对话协议 baseline。
Theory of Mind for Multi-Agent Collaboration via Large Language Models	2023	EMNLP 2023	important	theory-of-mind, belief-state, cooperative-game, collaboration	适合支撑“agent 之间需要理解彼此状态和意图”的协作机制章节。
Adaptive In-conversation Team Building for Language Model Agents	2024	arxiv	important	captainagent, team-building, agent-selection, in-conversation	补齐动态组队中“对话中途扩队/换队”的机制，可作为 AgentVerse、Dynamic Agent Network 之后的关键候选。
AgentCoord: Visually Exploring Coordination Strategy for LLM-based Multi-Agent Collaboration	2024	arxiv	important	coordination-strategy, visual-exploration, structured-representation, human-in-the-loop	适合支撑“协作协议/拓扑如何设计和调试”的工具化方向。
Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks	2024	arxiv	core	orchestrator, specialist-agents, replanning, autogenbench	是 orchestrator-specialist 架构的强参考系统，也提供 AutoGenBench 这类可隔离评测工具。
MedAide: Information Fusion and Anatomy of Medical Intents via LLM-based Agent Collaboration	2024	Information Fusion 2025	supporting	medical, intent-aware, role-rotation, information-fusion	应用较窄，但 intent-aware fusion 和 role rotation 对通用协作框架有借鉴价值。
Multi-Agent Collaboration in Incident Response with Large Language Models	2024	arxiv	important	incident-response, cybersecurity, centralized, decentralized, hybrid	可作为动态协作结构和高压任务响应的应用案例。
Scaling Large Language Model-based Multi-Agent Collaboration	2024	ICLR 2025 Poster	core	macnet, scaling-law, dag-topology, thousand-agents	是协作规模化和拓扑组织的重要论文，直接支撑“多个开源榜单 SOTA”和大规模协作框架目标。
Towards Effective GenAI Multi-Agent Collaboration: Design and Evaluation for Enterprise Applications	2024	arxiv	core	enterprise, coordination-mode, routing-mode, payload-reference	与项目落地到任务助手和企业级复杂任务高度贴合，尤其适合作端到端可用率与成本指标参考。
Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation	2025	AAAI 2026 Oral	core	arg-designer, autoregressive-graph-generation, topology-design, role-selection	直接补强自动组队与通信拓扑生成，是从固定团队走向按任务生成团队的高质量前沿工作。
Collaborating Action by Action: A Multi-agent LLM Framework for Embodied Reasoning	2025	arxiv	important	mindcraft, minecollab, embodied-agents, communication-efficiency	是“多 agent 通信并不总是越多越好”的重要负面证据，直接支撑通信压缩、意图路由和协作成本控制。
Collaborative Belief Reasoning with LLMs for Efficient Multi-Agent Collaboration	2025	arxiv	important	belief-modeling, intent-inference, embodied-collaboration, communication-efficiency	直接补强“协作者意图理解 + 低通信成本协作”，适合和 Theory of Mind、belief state 类工作放在同一协作脉络。
Mediator-Guided Multi-Agent Collaboration among Open-Source Models for Medical Decision-Making	2025	arxiv	supporting	mediator-agent, heterogeneous-models, medical-vqa, multimodal	适合借鉴 mediator-guided collaboration 和异构模型协作机制。
Talk Structurally, Act Hierarchically: A Collaborative Framework for LLM Multi-Agent Systems	2025	arxiv	core	structured-communication, hierarchical-refinement, talkhier, collaboration	直接补强“通信协议 + 分层协作”的设计路线，可作为动态团队协作的强 baseline。
Tiered Agentic Oversight: A Hierarchical Multi-Agent System for Healthcare Safety	2025	arxiv	important	hierarchical-oversight, safety, routing, healthcare	对项目的 hierarchical routing、error absorption、human-in-the-loop 协作有参考价值。
ConSensus: Efficient Multi-Agent Collaboration for Multimodal Sensing	2026	arxiv	important	multimodal, consensus, hybrid-fusion, token-efficient	是“专业 agent + 低成本融合”的好例子，适合任务助手多源信息整合场景。

Debate, Consensus & Verification

Paper	Year	Venue / source	Status	Tags	Why it matters
Can ChatGPT Defend its Belief in Truth? Evaluating LLM Reasoning via Debate	2023	Findings of EMNLP 2023	important	debate, reasoning-evaluation, belief-defense, verification	补充多 agent 辩论的评测视角，可用于分析共识形成是否真正提升事实性和推理。
ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate	2023	arxiv	important	multi-agent-evaluation, debate, llm-as-judge	可用于项目中的结果验证和 agent-as-judge 设计。
Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate	2023	EMNLP 2024	core	mad, divergent-thinking, judge, reasoning	是 debate/verification 协作主线的基础论文，已有条目保留为核心协作机制。
Improving Factuality and Reasoning in Language Models through Multiagent Debate	2023	arxiv	core	multiagent-debate, factuality, reasoning, society-of-minds	是 LLM multiagent debate 的代表性基础工作，应作为 debate 协作机制和 factuality 改进 baseline。
ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs	2023	ACL 2024	core	round-table, consensus, diverse-llms, confidence-weighted-voting	对异构 agent 的 consensus 机制、置信度字段和投票聚合策略很有参考价值。
GroupDebate: Enhancing the Efficiency of Multi-Agent Debate Using Group Discussion	2024	arxiv	core	group-debate, token-efficient, debate-groups	直接对应项目“协作不能增加 token/耗时”的成本约束。
Mixture-of-Agents Enhances Large Language Model Capabilities	2024	ICLR 2025 Spotlight	core	mixture-of-agents, layered-aggregation, multi-model-collaboration, ensemble	是多模型/多 agent 层级聚合的关键 baseline，也暴露了后续稀疏路由和成本控制需求。
More Agents Is All You Need	2024	TMLR 2024	important	agent-forest, sampling-and-voting, scaling-agents, ensemble	提供“更多 agent 是否值得”的基础 scaling 对照，可作为复杂协作机制的低成本 baseline。
Enhancing Multi-Agent Debate System Performance via Confidence Expression	2025	EMNLP 2025	important	confidence-expression, debate-dynamics, calibration	可用于设计 agent 消息协议中的 confidence 字段和动态仲裁机制。
iMAD: Intelligent Multi-Agent Debate for Efficient and Accurate LLM Inference	2025	AAAI 2026	core	selective-debate, trigger-policy, token-efficient, confidence	非常贴合“按需协作”和“成本不增”的产品目标。
Stay Focused: Problem Drift in Multi-Agent Debate	2025	EACL 2025	core	problem-drift, debate-failure, driftjudge, driftpolicy	重要负面结果，提醒协作轮次越多不一定越好，必须有终止和聚焦机制。
ARMOR-MAD: Adaptive Routing for Heterogeneous Multi-Agent Debate in Large Language Model Reasoning	2026	arxiv	core	adaptive-routing, heterogeneous-agents, debate	与项目的“能力波动、动态路由、成本控制”高度相关。
Courtroom-Style Multi-Agent Debate with Progressive RAG and Role-Switching for Controversial Claim Verification	2026	arxiv	important	courtroom-debate, progressive-rag, role-switching, claim-verification	对结构化 adversarial collaboration 和证据扩展机制有参考价值。
DynaDebate: Breaking Homogeneity in Multi-Agent Debate with Dynamic Path Generation	2026	arxiv	important	dynamic-path-generation, process-centric-debate, verification-agent	对“动态角色/路径分配 + verifier 触发机制”很有参考价值。
Multi-Agent Debate with Memory Masking	2026	arxiv	important	memory-masking, debate-robustness, erroneous-memory	直接支撑“关键信息传不准/错误信息传播”的上下文治理机制。
Tool-MAD: A Multi-Agent Debate Framework for Fact Verification with Diverse Tool Augmentation and Adaptive Retrieval	2026	arxiv	important	tool-augmented-debate, adaptive-retrieval, fact-verification	适合任务助手中的事实核验、工具专长协作和 evidence-aware debate。

Role-Playing, Agent Society & Simulation

Paper	Year	Venue / source	Status	Tags	Why it matters
CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society	2023	NeurIPS 2023	foundational	camel, role-playing, inception-prompting, agent-society	是 LLM 多 agent 角色协作和 agent society 的基础论文，应作为后续 CAMEL/role-playing 工作源头。
Generative Agents: Interactive Simulacra of Human Behavior	2023	UIST 2023	foundational	generative-agents, memory, reflection, social-simulation	对长期记忆、反思和群体涌现行为建模有基础价值，可支撑“协作系统如何积累经验”的背景。
SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents	2023	ICLR 2024 Spotlight	important	social-intelligence, role-play, interaction-evaluation, sotopia	可补足多 agent 协作评测中“社会智能、策略沟通、角色互动”这类非代码任务维度。
LLM Harmony: Multi-Agent Communication for Problem Solving	2024	arxiv	supporting	role-playing, personas, communication, problem-solving	可作为早期 role-playing collaboration baseline。

Software Engineering Multi-Agent Collaboration

Paper	Year	Venue / source	Status	Tags	Why it matters
ChatDev: Communicative Agents for Software Development	2023	ACL 2024	core	chatdev, chat-chain, communicative-dehallucination, software-development	是软件工程多 agent 协作的代表性工作，应作为固定流程和语言通信协作 baseline。
MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework	2023	arxiv	important	sop, role-specialization, software-agents	适合作固定 SOP 协作 baseline，用于对比动态拓扑和自进化方法。
AgentMesh: A Cooperative Multi-Agent Generative AI Framework for Software Development Automation	2025	arxiv	important	planner, coder, debugger, reviewer, software	是经典角色分工流水线 baseline，适合和动态/进化式协作比较。
Co-Saving: Resource Aware Multi-Agent Collaboration for Software Development	2025	arxiv	core	resource-aware, shortcuts, experiential-knowledge, token-reduction	直接服务项目“成本不增”和“从历史轨迹进化”的目标。
DocAgent: A Multi-Agent System for Automated Code Documentation Generation	2025	arxiv	important	documentation, topological-code-processing, verifier, orchestrator	适合参考“上下文拓扑排序 + 专业 agent + verifier”的协作结构。
Optimizing LLM-Based Multi-Agent System with Textual Feedback: A Case Study on Software Development	2025	arxiv	core	textual-feedback, prompt-optimization, group-optimization, software	是“根据失败解释优化多 agent 角色/prompt”的直接相关论文。
Shadows in the Code: Exploring the Risks and Defenses of LLM-based Multi-Agent Software Development Systems	2025	AAAI 2026	important	security, malicious-agent, software-development, defense	说明多 agent 协作系统必须有分角色安全和关键 agent 防护。
Towards Engineering Multi-Agent LLMs: A Protocol-Driven Approach	2025	APSEC 2025	core	protocol, structured-messaging, lifecycle, a2a, semap	与“通用协作框架”和 A2A/MCP 风格协议化设计高度贴合。
Self-Organizing Multi-Agent Systems for Continuous Software Development	2026	arxiv	core	self-organizing, continuous-development, manager-agents, hire-fire	强相关于动态 agent 组队、长期任务和自组织协作。

Robustness & Reliability

Paper	Year	Venue / source	Status	Tags	Why it matters
On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents	2024	ICML 2025 Poster	core	faulty-agents, resilience, challenge, reviewer	直接对应项目“agent 上下线、能力波动、整体任务执行不稳定”的痛点。
Amplified Vulnerabilities: Structured Jailbreak Attacks on LLM-based Multi-Agent Debate	2025	arxiv	important	jailbreak, multi-agent-debate, security, attack	说明多 agent 协作不天然更安全，必须把消息安全、角色安全和验证机制纳入设计。
Disagreements Can Help: A Self-Repair Approach to LLM Agents' Robustness in Knowledge Conflicts	2025	arxiv	important	knowledge-conflict, self-repair, disagreement	可用于设计冲突检测、互相质疑和自修复协作机制。
GUARDIAN: Safeguarding LLM Multi-Agent Collaborations with Temporal Graph Modeling	2025	arxiv	important	temporal-graph, safety, hallucination-propagation, anomaly-detection	直接服务于“关键信息传不准”和错误跨 agent 放大的安全问题，可作为 trace graph 防御方向。
Maris: A Formally Verifiable Privacy Policy Enforcement Paradigm for Multi-Agent Collaboration Systems	2025	arxiv	core	privacy-policy, message-flow-control, reference-monitor, prompt-injection	直接补强多 agent 通信的治理层，适合支撑敏感信息跨 agent 泄露和权限边界设计。
VerifyMAS: Hypothesis Verification for Failure Attribution in LLM Multi-Agent Systems	2026	arxiv	core	failure-attribution, trace-analysis, verification	自进化系统必须先知道失败来自分解、路由、执行、合成还是验证，故强相关。
When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems	2026	arxiv	important	safety, malicious-agent, message-defense	适合支撑多 agent 消息污染、防御和可信通信子方向。

Evolution & Optimization

Paper	Year	Venue / source	Status	Tags	Why it matters
Reflexion: Language Agents with Verbal Reinforcement Learning	2023	NeurIPS 2023	foundational	self-reflection, verbal-rl, episodic-memory	是自进化机制的单 agent 基础，可扩展到团队级经验沉淀。
Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation	2023	arxiv	foundational	self-improvement, scaffold-optimization, code-generation	可作为自动改 prompt/tool/workflow 的思想背景。
Voyager: An Open-Ended Embodied Agent with Large Language Models	2023	TMLR 2024	foundational	lifelong-learning, skill-library, self-verification	支撑“技能库/经验库”式长期进化，但本身不是多 agent 主线。
AFlow: Automating Agentic Workflow Generation	2024	ICLR 2025 Oral	core	workflow-search, mcts, agentic-workflow	适合作离线 workflow 自动优化器，对项目“可自进化协作系统”非常关键。
EvoAgent: Towards Automatic Multi-Agent Generation via Evolutionary Algorithms	2024	NAACL 2025	core	evolutionary-algorithm, multi-agent-generation, mutation, crossover	与项目“群体进化”命名和目标直接对齐。
ExpeL: LLM Agents Are Experiential Learners	2024	AAAI 2024	important	experiential-learning, memory, reflection, agent-learning	虽然不是多 agent 专用，但为“从执行轨迹中积累经验并改进行为”的进化闭环提供关键基线。
Experiential Co-Learning of Software-Developing Agents	2024	ACL 2024	important	experiential-learning, software-agents, co-learning, trajectory-memory	直接连接多 agent 软件协作和经验驱动进化，可作为 ChatDev 类系统如何从失败和轨迹中持续改进的核心参考。
GPTSwarm: Language Agents as Optimizable Graphs	2024	ICML 2024 Oral	core	gptswarm, graph-optimization, prompt-optimization, edge-optimization	几乎直接对应“多 agent 反馈链路如何优化”，可作为群体进化主 baseline。
AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems	2025	arxiv	important	decentralized-coordination, agent-evolution, rag, dynamic-routing	补齐去中心化协作进化路线，与 centralized orchestrator/manager 架构形成重要对照。
Cross-Task Experiential Learning on LLM-based Multi-Agent Collaboration	2025	arxiv	core	mael, cross-task-experience, graph-collaboration, reward-pool	直接补强“跨任务经验积累”和“协作系统进化”方向。
MAS-GPT: Training LLMs to Build LLM-based Multi-Agent Systems	2025	ICML 2025 Poster	core	mas-generation, query-adaptive, executable-code, system-design	直接面向“根据任务自动生成协作系统”的进化方向，可与 AFlow、SwarmAgentic、ARG-Designer 共同构成自动化设计主线。
Multi-Agent Collaboration via Evolving Orchestration	2025	NeurIPS 2025	core	evolving-orchestration, puppeteer, reinforcement-learning, dynamic-sequencing	直接补齐“协作编排本身可进化”的顶会工作，适合和 GPTSwarm、AFlow、EvoAgent、EvoMAS 共同作为进化主线 baseline。
SwarmAgentic: Towards Fully Automated Agentic System Generation via Swarm Intelligence	2025	EMNLP 2025	important	agentic-system-generation, swarm-intelligence, automated-design, evolution	直接补充多 agent 系统自动生成与进化方向，适合与 AFlow/GPTSwarm/EvoAgent 放在同一脉络。
Evolutionary Generation of Multi-Agent Systems	2026	arxiv	core	evomas, configuration-space, evolutionary-generation	对工程落地尤其重要，因为配置空间进化比任意代码进化更可控。
Evolve as a Team: Collaborative Self-Evolution for LLM-based Multi-Agent Systems	2026	arxiv	core	team-evolution, collaborative-self-evolution, post-task-communication	最贴合“多 agent 之间互相影响的复杂反馈链路”的研究问题。
Swarm Skills: A Portable, Self-Evolving Multi-Agent System Specification	2026	arxiv	core	swarm-skills, portable-specification, self-evolving	对“通用协作框架、可迁移到其他智能体产品线”非常贴合。

Evaluation & Benchmarks

Paper	Year	Venue / source	Status	Tags	Why it matters
AgentBench: Evaluating LLMs as Agents	2023	ICLR 2024 Poster	important	agent-benchmark, interactive-environment, evaluation	可作为任务助手多 agent 系统的公共能力评测参考。
GAIA: a benchmark for General AI Assistants	2023	ICLR 2024 Poster	important	general-assistant, web, tool-use, benchmark	适合作“多步骤信息搜集+整合”公共评测集。
SWE-bench: Can Language Models Resolve Real-World GitHub Issues?	2023	ICLR 2024 Oral	important	software-engineering, long-horizon, benchmark	适合测长上下文、多步骤协作和可恢复执行，但需补充非代码长尾任务。
Agent-as-a-Judge: Evaluate Agents with Agents	2024	ICML 2025 Poster	core	agent-evaluation, process-feedback, llm-judge	自进化需要 reward/critic，该论文可支撑反馈信号设计。
VillagerAgent: A Graph-Based Multi-Agent Framework for Coordinating Complex Task Dependencies in Minecraft	2024	arxiv	important	villagerbench, minecraft, dag-coordination, complex-dependencies	适合作为复杂依赖、多 agent 分工、同步和动态适应能力的协作评测场景。
MALLM: Multi-Agent Large Language Models Framework	2025	EMNLP 2025 Demo	important	multi-agent-debate, framework, configuration, evaluation-pipeline	适合系统比较 debate、relay、memory、voting、consensus 等多 agent 策略，补齐“实验框架/可复现评测”维度。
MultiAgentBench: Evaluating the Collaboration and Competition of LLM agents	2025	ACL 2025	core	benchmark, collaboration, competition, coordination-protocols	是当前多 agent 协作评测的重要 ACL 基准，可用于验证通信拓扑和协作策略是否真正提升任务完成与协作质量。
MARBLE: Multi-Agent Reasoning for Bioinformatics Learning and Evolution	2026	arxiv	supporting	bioinformatics, multi-agent-reasoning, learning, evolution	虽是垂直领域，但对“领域复杂任务 + 多 agent 学习/进化”有参考价值，适合作补充性 benchmark。

Surveys & Taxonomies

Paper	Year	Venue / source	Status	Tags	Why it matters
A Survey of Multi-Agent Deep Reinforcement Learning with Communication	2024	Autonomous Agents and Multi-Agent Systems 2024	important	survey, marl, communication, protocols	为 LLM 多 agent 通信研究提供前 LLM 时代的系统背景，帮助区分学到的通信协议与提示式通信。
The Landscape of Emerging AI Agent Architectures for Reasoning, Planning, and Tool Calling	2024	arxiv	survey	survey, agent-architecture, planning, tool-use	适合作 agent architecture 背景章节。
Multi-Agent Collaboration Mechanisms: A Survey of LLMs	2025	arxiv	survey	survey, collaboration, coordination-protocols	适合作本综述的分类法参考。
Multi-Agent Coordination across Diverse Applications: A Survey	2025	arxiv	survey	survey, coordination, scalability, heterogeneity	适合支撑“协作”的理论定义和跨领域背景。

GitHub Project Catalog

Frameworks & Runtimes

Project	Status	Tags	Why it matters
AG2	important	autogen-fork, agentos, groupchat, multi-agent	适合与 Microsoft AutoGen 对比，观察社区版多 agent 框架如何演进。
Agent Squad	important	routing, multi-agent, specialist-agents, context	与项目“任务助手多专业 agent 联合执行”和 request routing 高度相关。
AgentScope	core	multi-agent-platform, message-exchange, fault-tolerance, distributed	与“通用协作框架 + 动态协作 + 鲁棒部署”高度贴合。
AgentVerse	important	multi-agent, framework, simulation, collaboration	可作为早期通用多 agent 协作框架和仿真实验平台参考。
Agno	important	agent-platform, teams, memory, runtime	适合参考多 agent team 在生产平台中的运行、管理和观测抽象。
AutoGen	core	multi-agent, event-driven, agentchat, runtime	适合参考多 agent conversation、事件驱动通信和分布式 agent runtime。
BeeAI Framework	important	python, typescript, workflows, production-agents	可作为企业级多 agent 运行时、工具接入和协议生态的补充参考。
CAMEL	core	agent-society, multi-agent, memory, benchmarks	适合做多 agent 实验平台和 agent society baseline。
ChatDev	core	software-company, chat-chain, multi-agent, macnet	是软件工程多 agent 协作和协作 scaling law 的核心参考实现。
CrewAI	important	crews, flows, role-based, business-workflows	可作为业务流程协作 baseline，但研究创新不应停留在框架调用。
deepagents	important	long-running, subagents, planning, langgraph	适合参考长尾复杂任务中 subagent 编排、任务持久化和工作状态管理。
Google ADK	important	agent-development-kit, multi-agent, evaluation, deployment	可作为多 agent 产品化 SDK、评测和部署路径的强参考。
LangGraph	core	graph-runtime, stateful, durable-execution, multi-agent	适合做多 agent 协作运行时和 traceable workflow baseline。
LlamaIndex	important	agent-workflows, retrieval, tools, multi-agent	对“多 agent + 检索/文档/数据任务”的任务助手落地很重要，适合作为 RAG-heavy 协作基线。
Mastra	important	typescript, workflows, agents, memory	对前端/Node 生态中多 agent workflow 与产品化部署很有参考价值。
MetaGPT	important	sop, role-specialization, software-company	适合做固定角色/固定流程协作 baseline。
Microsoft Agent Framework	important	multi-agent, workflows, orchestration, observability	适合跟踪 AutoGen 之后的生产级多 agent 编排方向。
NeMo Agent Toolkit	important	agent-toolkit, multi-agent, optimization, nvidia	与“多 agent 协作优化”和企业级 agent team 运行效率直接相关。
Open Multi-Agent	important	dag, tracing, typescript, multi-agent	适合作任务 DAG、并行协作和可观测执行 baseline。
OpenAI Agents SDK	important	agents, handoffs, tracing, guardrails, mcp	适合借鉴 handoff、manager-style orchestration 和 tracing 的简洁抽象。
OpenAI Swarm	important	educational, handoffs, routines, multi-agent-orchestration	虽非生产框架，但对理解 handoff、agent-as-tool 和轻量编排模式非常有参考价值。
PraisonAI	important	multi-agent, workforce, self-improving, rag	可作为业务流程中“多角色 agent workforce”落地方式的对照项目。
Pydantic AI	important	typed-agents, structured-output, graph, production	多 agent 协作需要强 schema、可验证状态和清晰工具契约，Pydantic AI 是重要工程参考。
Semantic Kernel	important	enterprise, orchestration, plugins, multi-agent	适合作企业场景中多 agent 编排、工具接入和结构化插件契约的工程参考。
smolagents	important	code-agents, managed-agents, tools, lightweight	对“简洁 agent loop + managed subagents”的最小实现有参考价值，可作为复杂框架的低开销对照。
Strands Agents	important	sdk, multi-agent, mcp, production	对“通用协作框架 + 跨模型/跨工具生态”有参考价值。

Protocols & Interoperability

Project	Status	Tags	Why it matters
Agent Client Protocol	important	acp, editor-agent, protocol, sessions	对跨宿主 agent 协作、agent handoff 和可迁移会话协议有参考价值。
Agent Communication Protocol	supporting	acp, agent-protocol, interoperability, archived	与 A2A、ANP、MCP 一起构成跨 agent 通信协议调研边界，应在协议对比中说明其归档状态。
Agent Network Protocol	important	anp, agent-network, interoperability, decentralized	补齐 A2A/MCP 之外的协议路线，适合研究跨组织 agent 发现和协作网络。
Agent2Agent Protocol	core	a2a, agent-interop, task-delegation, agent-cards	与“通用协作框架”和跨产品线 agent 协作高度相关。
Agora Protocol	supporting	agora, protocol, python-sdk, collaboration	可作为 ProtocolBench 类协议评测中的协议生态补充，不作为核心工程框架。
mcp-agent	important	mcp, workflow, multi-server, agent-framework	适合参考如何把多个工具服务器聚合进 agent 协作流程。
Model Context Protocol	core	mcp, tools, resources, prompts, interoperability	多智能体框架若要跨产品线复用，工具/上下文接入层建议兼容 MCP。

Benchmarks & Evaluation

Project	Status	Tags	Why it matters
AgentBench	important	agent-benchmark, environments, reasoning, decision-making	适合测基础 agent 执行能力，但需补充多 agent 通信和进化指标。
ChatEval	important	multi-agent-evaluator, debate, llm-as-judge	可作为多 agent verifier/judge 组件和评价方法 baseline。
Collaborative Gym	important	collaborative-agents, human-agent-collaboration, evaluation, environments	补齐 human-agent/team collaboration 评测维度，有助于评估多 agent 与人类协同的真实可用性。
DeepEval	supporting	evaluation, llm-judge, testing	可作为自建 Qianwen-LongTail-MAS-Bench 的评测基础设施参考。
E2EDev	important	end-to-end-software-development, bdd, multi-agent-annotation, benchmark	适合评估多 agent 从需求到交付的完整长任务能力。
M-MAD	supporting	multi-agent-debate, machine-translation-evaluation, llm-as-judge	适合参考如何把复杂评价标准拆成多个 specialist judges。
MARBLE / MultiAgentBench	core	multiagentbench, marble, acl-2025, collaboration, competition	对验证多 agent 通信拓扑、协作协议和竞争/协作行为质量非常关键。
Multi-Agents-Debate	core	multi-agent-debate, reasoning, judge	是 debate 协作机制的核心 baseline。
Promptfoo	supporting	eval, red-team, ci, regression	适合做多 agent 变更后的回归测试和安全测试。
SWE-bench	important	software-engineering, long-horizon, benchmark	可测长任务执行、上下文管理、验证和恢复能力。
SWE-Dev	important	feature-development, software-engineering, benchmark, rl	适合补充 SWE-bench 对“新增功能/端到端开发”的覆盖。
tau2-bench	important	tool-agent-user, benchmark, multi-turn, real-world-domains	适合评估多 agent 任务助手在真实工具和用户反馈循环中的协作表现。
Terminal-Bench	important	terminal, benchmark, long-horizon, verification	可用于测试多 agent 在代码、命令行、文件系统任务中的长程协作和恢复能力。
TheAgentCompany	important	workplace, simulated-company, benchmark, long-horizon	与多 agent 任务助手和软件团队协作评测高度相关。
WebArena	important	web-agent, benchmark, realistic-environment, tool-use	可作为任务助手“多步骤信息搜集+执行”的公共环境基准。
WorkArena	important	enterprise, web-agent, service-workflows, benchmark	适合补充企业长尾任务、流程系统和真实业务操作场景的评测。

Observability & Operations

Project	Status	Tags	Why it matters
AgentOps	important	agent-monitoring, tracing, cost, benchmarks	多 agent 通信、成本和失败归因需要统一轨迹与成本观测，AgentOps 是强工程参考。
Arize Phoenix	supporting	observability, eval, tracing, llmops	可作为 trace + evaluation + failure analysis 工程参考。
GUARDIAN	important	temporal-graph, safety, anomaly-detection, multi-agent	适合把多 agent 消息 trace 转成图并做错误传播/异常检测。
Langfuse	supporting	observability, traces, cost, prompts	适合记录多 agent 消息、成本、失败轨迹和进化前后对比。
LangSmith	supporting	observability, tracing, eval, debugging	多 agent 通信/进化需要 trace-native evaluation，LangSmith 可作工程参考。
LangWatch	important	evaluations, simulations, agent-testing, observability	适合作多 agent 变更后的离线仿真、回归评测和线上质量监控。
Weave	supporting	tracing, evaluations, llmops, wandb	可用于多 agent 轨迹回放、成本/质量对比和进化前后评估。

Research Implementations & Baselines

Project	Status	Tags	Why it matters
AgentForest	important	sampling-and-voting, agent-scaling, ensemble-baseline	适合作为最简单但必须比较的多 agent scaling baseline。
AgentNet	important	decentralized, evolutionary-coordination, rag, dynamic-routing	与中心化 orchestrator 架构形成对照，适合研究去中心化协作和进化机制。
AgentPrune	core	communication-pruning, token-efficiency, iclr-2025, robustness	是低成本通信和通信鲁棒性实验的重要 baseline，适合直接纳入复现实验。
ARG-Designer	core	aaai-2026, topology-generation, role-selection, graph-generation	对自动组队、自动拓扑设计和低成本协作非常关键，适合作为前沿 baseline。
Dr. MAS	important	reinforcement-learning, co-training, heterogeneous-llms, multi-agent	对群体进化和多 agent 学习机制有参考价值，适合作为训练型 MAS baseline。
Generative Agents	important	agent-society, memory, reflection, simulation	适合作为长期记忆、经验沉淀和 agent society 章节的经典参考实现。
GPTSwarm	core	optimizable-graphs, prompt-optimization, edge-optimization, evolution	是群体进化和“多 agent 反馈链路优化”的核心实现，应作为主 baseline。
MALLM	important	multi-agent-llms, consensus, debate, framework	适合快速搭建 debate、投票、共识类 baseline，并记录多 agent 实验配置。
MARTI	important	reinforced-training, inference, multi-agent-llm, optimization	补齐“多 agent 不只推理编排，也可训练/强化优化”的工程路线。
MAS-GPT	core	mas-generation, icml-2025, executable-code, auto-design	是自动化多 agent 系统生成和进化方向的重要实现。
MasRouter	core	routing, acl-2025, llm-selection, multi-agent	是动态路由、低成本协作和模型选择实验的核心 baseline。
RADAR	supporting	communication-structure, diffusion, redundancy-aware, topology-generation	作为 2026 通信结构生成候选 baseline，可与 G-Designer、ARG-Designer 对比。
ReConcile	core	round-table, consensus, confidence-weighted-voting, diverse-models	可作为异构 agent 共识、置信度表达和投票聚合的基础实现。
RMoA	important	residual-mixture-of-agents, diversity-selection, adaptive-termination	可作为 MoA 家族中“信息保真 + 成本控制”的重要对照。
SMoA	important	sparse-mixture-of-agents, response-selection, early-stopping	适合研究多 agent 聚合如何在保持效果的同时降低成本。
SOTOPIA	important	social-intelligence, role-play, interaction-evaluation	可补充非代码长尾任务中的协作、沟通和社会目标评测。
TalkHier	core	structured-communication, hierarchical-refinement, llm-multi-agent	适合做“结构化消息协议 + 分层协作修正”的核心 baseline。
VillagerAgent	important	minecraft, villagerbench, dag-coordination, complex-dependencies	适合构造任务依赖、同步、动态适应和 hallucination 控制的协作评测。

Reference Systems & Case Studies

Project	Status	Tags	Why it matters
DocAgent	important	documentation, verifier, orchestrator, code-understanding	是“专业 agent + topological context + verifier”的清晰参考系统。
Magentic-One	core	orchestrator, specialist-agents, autogenbench, replanning	是 orchestrator-subagent 架构、错误恢复和隔离评测的高价值参考系统。
OpenHands	important	software-agent, coding, benchmark, runtime	虽不是纯多 agent 框架，但可作为长任务执行、工具使用和软件任务评测的重要 baseline。
SWE-agent	important	software-agent, swe-bench, execution, repair	可作为单/多 agent 软件任务对比基线，帮助衡量多 agent 协作是否真正带来收益。
Symphony	important	orchestration, issue-driven, isolated-runs, coding-agents	虽偏编码 agent，但对“任务分解、隔离执行、编排控制面和回收结果”非常有参考价值。

Technical Docs, Blogs, and Field Reports

Anthropic Official Engineering Articles

Reading	Status	Tags	Why it matters
Anthropic - Building Effective AI Agents	core	engineering-guide, workflows, agents, patterns	是多 agent 协作模式选择的基础阅读，能防止把所有任务都错误地做成复杂多 agent。
Anthropic - Building multi-agent systems	core	field-report, context-protection, specialization, token-cost	直接约束本项目“token 消耗和端到端耗时维持不增”的工程边界。
Anthropic - Claude Code auto mode	supporting	permissions, autonomy, classifier, coding-agents	对多 agent 中 human-in-the-loop、权限门控和自动化边界设计有启发。
Anthropic - Claude Code sandboxing	supporting	sandboxing, security, autonomy, coding-agents	多 agent 协作会放大工具调用和文件系统风险，沙箱边界是基础治理机制。
Anthropic - Code execution with MCP	important	mcp, code-execution, sandbox, tool-boundaries	多 agent 系统需要清晰工具边界和执行隔离，MCP 是关键互操作层。
Anthropic - Demystifying evals for AI agents	important	agent-evals, nondeterminism, trajectories, evaluation	自进化多 agent 必须有可靠 eval/reward，这篇补齐评测方法论。
Anthropic - Effective context engineering for AI agents	core	context-engineering, memory, working-state, token-budget	多 agent 通信路由本质上依赖上下文预算和信息选择，这篇可直接支撑通信层设计。
Anthropic - Effective harnesses for long-running agents	important	long-running-agents, state, resumability, harness	任务助手长尾复杂任务通常是长时任务，多 agent 协作运行时必须处理恢复和状态漂移。
Anthropic - Harness design for long-running application development	important	long-running, app-development, harness, reliability	可为软件工程多 agent 协作和持续开发任务提供工程约束。
Anthropic - How we built our multi-agent research system	core	field-report, lead-agent, parallel-subagents, research	是“多步骤信息搜集+整合”场景最直接的公开工程案例，应作为系统设计和成本评估的必读材料。
Anthropic - Multi-agent coordination patterns	core	patterns, generator-verifier, orchestrator-subagent, message-bus, shared-state	可作为本仓库协作分类和系统架构章节的工程模式骨架。
Anthropic - Quantifying infrastructure noise in agentic coding evals	supporting	evaluation, infrastructure-noise, coding-agents, reproducibility	多 agent 系统评测更容易受环境和基础设施波动影响，这篇适合指导评测治理。
Anthropic - Scaling Managed Agents	important	managed-agents, session-logs, sandbox, architecture	对多 agent 运行时边界、隔离执行和状态回放设计很有参考价值。
Anthropic - Writing effective tools for AI agents	important	tool-design, schemas, agent-tools, reliability	多 agent 框架中的 specialist agents 常以工具形式互相调用，工具契约质量直接影响协作可靠性。

OpenAI Official Engineering Articles and Docs

Reading	Status	Tags	Why it matters
OpenAI - Building more helpful agents with a new evaluation framework	important	field-report, browsecomp, evaluation, hard-to-find-information	适合支撑“多步骤信息搜集+整合”类任务助手评测，尤其是高难检索和证据链质量。
OpenAI - Harness engineering	core	harness-engineering, verification, reliability, agent-first-software	为多 agent 工程化提供 harness、验证和可靠性约束，是系统落地的重要方法论。
OpenAI - Symphony orchestration	important	orchestration, codex, issue-driven, control-plane	可借鉴到多 agent 任务助手的任务队列、隔离执行和结果合并流程。
OpenAI - The next evolution of the Agents SDK	important	agents-sdk, sandbox, manifests, memory, tools	对生产级 agent SDK 如何标准化工具、记忆和执行环境有参考价值。
OpenAI - Unrolling the Codex agent loop	important	agent-loop, tool-replay, stateless-calls, codex	多 agent 系统同样需要处理轨迹回放、上下文增长和工具调用状态，这篇能支撑运行时设计。
OpenAI Agents SDK - Orchestrating Multiple Agents	core	documentation, agents-sdk, handoffs, orchestration	对用 OpenAI Agents SDK 复现 supervisor/manager/handoff 协作模式非常关键。
OpenAI Cookbook - Orchestrating Agents	important	cookbook, handoffs, routines, orchestration	是 OpenAI 生态中最直接的多 agent 编排实践示例，应和 Swarm/Agents SDK 一起参考。
OpenAI Developers - Testing Agent Skills Systematically with Evals	important	skills, evals, traces, regression	自进化多 agent 系统需要 regression-safe 的技能/协作模式评测闭环。

Multi-Agent Framework Documentation

Reading	Status	Tags	Why it matters
AG2 Documentation - Multi-agent Chat	supporting	documentation, ag2, groupchat, multi-agent	适合与 AutoGen 官方模式对照，观察 group chat 协作抽象的演进。
AutoGen Documentation - Multi-agent Design Patterns	important	documentation, autogen, design-patterns, runtime	适合把 AutoGen 论文和 Magentic-One 系统落到可复现实验配置。
Hugging Face smolagents - Multi-agent Systems	important	documentation, smolagents, managed-agents, code-agents	适合用最小 agent 框架验证“manager + specialist”的协作开销和效果。
LangGraph Documentation - Multi-agent Systems	important	documentation, langgraph, supervisor, swarm	是构建 traceable multi-agent workflow baseline 的直接工程参考。
LlamaIndex Documentation - Multi-agent Workflows	important	documentation, llamaindex, workflows, retrieval-agents	对多 agent 任务助手中的数据检索、工具调用和 workflow 组合有直接工程价值。
Mastra Documentation - Agent Networks	important	documentation, mastra, agent-network, typescript	补齐 Node/TypeScript 生态下多 agent network 工程实践。
Pydantic AI Documentation - Multi-agent Applications	important	documentation, pydantic-ai, typed-agents, applications	对要求高可靠 schema、工具参数和状态校验的多 agent 系统尤其有价值。

Protocol Documentation

Reading	Status	Tags	Why it matters
Google A2A Protocol Documentation	important	documentation, a2a, interoperability, agent-cards	对“通用协作框架”和跨产品线 agent 协作协议设计高度相关。
Model Context Protocol Documentation	important	documentation, mcp, tools, context	多 agent 系统若要可迁移和可扩展，工具/上下文层应优先兼容 MCP。
ProtocolBench - Which LLM Multi-Agent Protocol to Choose?	supporting	protocolbench, protocolrouter, a2a, anp, acp, mcp	虽未作为正式顶会论文收录，但对协议选型、协议评测指标和 hybrid protocol routing 很有参考价值。

Evaluation, Runtime, and Engineering Blogs

Reading	Status	Tags	Why it matters
Cognition - Don't Build Multi-Agents	important	anti-patterns, agent-design, reliability, single-agent-baseline	是重要负面/约束性阅读，可帮助判断什么时候不该上多 agent，防止复杂度和成本无谓膨胀。
Cognition - What We Learned Building Cloud Agents	important	cloud-agents, vm-isolation, snapshots, governance	对多 agent 长任务运行环境、隔离执行和企业落地非常有参考价值。
HumanLayer - 12 Factor Agents	important	production-agents, principles, operations, reliability	适合作多 agent 任务助手工程化 checklist。
HumanLayer - Skill Issue: Harness Engineering for Coding Agents	supporting	harness-engineering, coding-agents, skills, operations	可迁移到多 agent 中 skill/workflow/tool 描述的系统化管理。
Inngest - Your Agent Needs a Harness, Not a Framework	important	harness-first, reliability, orchestration, production	对多 agent 系统的队列、重试、状态和事件驱动 orchestration 很有启发。
LangChain - Agent frameworks, runtimes, and harnesses	important	architecture, framework, runtime, harness	有助于把多 agent 通信/协作算法和底层 harness 工程职责分离。
LangChain - Evaluating Deep Agents	important	deep-agents, evaluation, long-horizon, stateful-agents	适合设计多 agent 长尾任务评测、错误分类和回归测试。
LangChain - How and when to build multi-agent systems	important	engineering-guide, supervisor, handoff, multi-agent	可帮助把论文中的协作模式映射到 LangGraph/LangChain 工程实现。
LangChain - Improving Deep Agents with harness engineering	important	deep-agents, harness-engineering, benchmark, iteration	支撑“自进化不一定先改模型，可以先优化 harness/workflow”的研究路线。
LangChain - The Anatomy of an Agent Harness	important	harness, architecture, components, reliability	是把多 agent 框架落成可维护工程系统的架构参考。
Martin Fowler - Harness Engineering	important	architecture, harness-engineering, entropy, gen-ai	为多 agent 系统边界、治理和工程复杂度控制提供高质量架构视角。

High-Quality Personal and Community Blogs

Reading	Status	Tags	Why it matters
DeepLearning.AI - Four AI Agent Strategies	important	agentic-design-patterns, multi-agent-collaboration, reflection, planning	是工业界传播最广的 agentic pattern 框架之一，适合放在多 agent 协作模式章节作为概念入口。
LessWrong - Survey of Multi-agent LLM Evaluations	supporting	multi-agent-evaluation, risks, miscoordination, collusion	补齐多 agent 评测不只看任务分数，还要看协作失败、串谋和安全风险的视角。
Lilian Weng - LLM Powered Autonomous Agents	important	agent-overview, planning, memory, tool-use, reflection	虽不是多 agent 专文，但为理解多 agent 中每个 agent 的能力边界和自我改进机制提供基础。
Simon Willison - Agent Definition	supporting	agent-definition, tool-loop, terminology, multi-agent	有助于统一“agent”和“multi-agent system”的工程术语，避免把普通 workflow 误称为多 agent。

Non-GitHub Project and Benchmark References

Protocols & Interoperability

Reference	Status	Tags	Why it matters
SEMAP	core	protocol, structured-messaging, a2a, software-engineering	虽主要是论文方案，但非常适合作为多 agent 协作协议设计参考。

Benchmarks & Evaluation

Reference	Status	Tags	Why it matters
GAIA	important	general-assistant, web, tool-use, benchmark	适合任务助手信息搜集、整合和工具使用评测。

Observability & Operations

Reference	Status	Tags	Why it matters
AgentCoord	supporting	coordination-visualization, strategy-design, human-in-the-loop	适合作协作拓扑、任务依赖和 agent 关系的调试工具参考。

Research Implementations & Baselines

Reference	Status	Tags	Why it matters
MINDCraft / MineCollab	important	minecraft, communication-efficiency, embodied-agents, collaboration-benchmark	可作为 embodied multi-agent 场景下“通信越多是否越好”和“协作是否真正带来收益”的强评测补充。

Maintenance

Source of truth:

data/papers.yaml
data/projects.yaml

Regenerate generated files:

python3 scripts/render.py

Validate schema and links:

python3 scripts/verify.py

Verify GitHub-backed project entries with GitHub CLI:

python3 scripts/verify_github_projects.py

Contributors

Flashinglf

1 commits

Similar repos

curation-labs/awesome-agent-coordination

kyegomez/awesome-multi-agent-papers

A compilation of the best multi-agent papers

1,689

vchaitanyachowdari/awesome-multi-agent-papers

A Compilation of the best Multi Agent Papers

luo-junyu/Awesome-Agent-Papers

[Up-to-date] Large Language Model Agent: A Survey on Methodology, Applications and Challenges

2,862

Yigtwxx/awesome-claude-multi-agent

A curated list of frameworks, patterns, and research for orchestrating multi-agent systems built on…

johnsutor/awesome-multi-agent-optimization

A collection of papers and libraries for performing multi-agent optimization

EvoMap/awesome-agent-swarm

A curated list of AI Agent Swarm frameworks, multi-agent orchestration, swarm intelligence, and…

Andrewzh112/Awesome-LLM-based-MultiAgents

Languages

Python

100.0%

Flashinglf/awesome-multi-agent-communication-collaboration-evolution

Python

1 commits

updated Jun 24, 2026

See the code

README

Awesome Multi-Agent Communication, Collaboration, and Evolution

A curated and updateable repository for research on multi-agent communication, collaboration, robustness, and evolution, focused on LLM-based multi-agent systems and long-tail task assistants.

Papers: 108
GitHub projects and systems: 79
Technical docs, blogs, and field reports: 47
Non-GitHub project/reference rows: 4
Core papers: 42
Core GitHub projects/systems: 17
Core technical readings: 7
Core non-GitHub references: 1
Papers with non-arXiv primary source: 41
Papers with verified venue metadata: 53
Last verified: 2026-06-13
Language: English | 中文

Featured Technical Readings

Anthropic - How we built our multi-agent research system: Anthropic 对生产级 research multi-agent system 的工程复盘，包含 lead agent、并行 subagents、任务分解、trace 和 citation 验证经验。
Anthropic - Building multi-agent systems: 讨论什么时候值得使用多 agent，以及 context protection、parallelization、specialization 三类收益和高额 token 成本之间的取舍。
Anthropic - Multi-agent coordination patterns: 总结 generator-verifier、orchestrator-subagent、agent teams、message bus、shared state 五类常见协作模式及适用条件。
Anthropic - Building Effective AI Agents: Anthropic 对 workflows 与 autonomous agents 的实用区分，并总结 prompt chaining、routing、parallelization、orchestrator-workers、evaluator-optimizer 等模式。
OpenAI Agents SDK - Orchestrating Multiple Agents: OpenAI Agents SDK 官方多 agent 编排文档，覆盖 agents-as-tools、handoffs 和 manager 模式。
OpenAI Cookbook - Orchestrating Agents: OpenAI Cookbook 示例，展示如何用 routines、handoffs 和工具组织多个 agents。
OpenAI - Harness engineering: OpenAI 对 harness engineering 的实践总结，强调通过约束、验证和工具环境让 agent-first 软件更可靠。
OpenAI - Building more helpful agents with a new evaluation framework: OpenAI 发布 BrowseComp，用难以搜索定位的问题评估 agent 深度检索和信息整合能力。
LangChain - How and when to build multi-agent systems: LangChain 对 multi-agent supervisor、handoff、subagent 工具化等模式的实践建议。
AutoGen Documentation - Multi-agent Design Patterns: AutoGen 官方文档中的多 agent 设计模式入口，覆盖 Core runtime 下的协作组织方式。

Scope
Featured Technical Readings
Main Documents
Category Overview
Source Coverage
Core Papers
Core GitHub Projects and Systems
Core Technical Readings
Core Non-GitHub References
Complete Paper Catalog
GitHub Project Catalog
Technical Docs, Blogs, and Field Reports
Non-GitHub Project and Benchmark References
Maintenance

Scope

This repository is intentionally narrower than generic agent-framework lists. It tracks work that helps answer three questions:

Communication: who should receive what information, when, and under what token budget?
Collaboration: how should agents form teams, coordinate, verify, and recover from failures?
Evolution: how can traces and feedback improve prompts, roles, tools, edges, topologies, and workflows?

Main Documents

Category Overview

Catalog	Category	Entries
Papers	Communication & Routing	22
Papers	Task-Oriented Collaboration & Dynamic Teaming	22
Papers	Debate, Consensus & Verification	16
Papers	Role-Playing, Agent Society & Simulation	4
Papers	Software Engineering Multi-Agent Collaboration	9
Papers	Robustness & Reliability	7
Papers	Evolution & Optimization	16
Papers	Evaluation & Benchmarks	8
Papers	Surveys & Taxonomies	4
GitHub projects	Frameworks & Runtimes	26
GitHub projects	Protocols & Interoperability	7
GitHub projects	Benchmarks & Evaluation	16
GitHub projects	Observability & Operations	7
GitHub projects	Research Implementations & Baselines	18
GitHub projects	Reference Systems & Case Studies	5
Technical docs/blogs	Anthropic Official Engineering Articles	14
Technical docs/blogs	OpenAI Official Engineering Articles and Docs	8
Technical docs/blogs	Multi-Agent Framework Documentation	7
Technical docs/blogs	Protocol Documentation	3
Technical docs/blogs	Evaluation, Runtime, and Engineering Blogs	11
Technical docs/blogs	High-Quality Personal and Community Blogs	4
Non-GitHub references	Protocols & Interoperability	1
Non-GitHub references	Benchmarks & Evaluation	1
Non-GitHub references	Observability & Operations	1
Non-GitHub references	Research Implementations & Baselines	1

Source Coverage

Primary source	Papers
arxiv	67
openreview	16
acl	10
neurips	4
pmlr	4
aaai	2
journal	2
acm	1
ieee	1
ijcai	1

Core Papers

Area	Paper	Year	Venue / source	Why it matters
Communication & Routing	RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory	2025	arxiv	直接命中项目“通信代价大、关键信息传不准”的问题，可作为预算约束上下文路由主 baseline。
Communication & Routing	Improving Multi-Agent Debate with Sparse Communication Topology	2024	arxiv	支撑“不是所有 agent 都应该互相通信”的设计假设，适合做全连接通信的对照实验。
Communication & Routing	Dynamic Generation of Multi-LLM Agents Communication Topologies with Graph Diffusion Models	2025	arxiv	适合支撑“动态 agent 组队 + 动态通信网络”的算法路线。
Communication & Routing	Towards Adaptive, Scalable, and Robust Coordination of LLM Agents: A Dynamic Ad-Hoc Networking Perspective	2026	arxiv	同时覆盖通信路由、动态协作和鲁棒性，是项目基础设施设计的强相关论文。
Task-Oriented Collaboration & Dynamic Teaming	A Dynamic LLM-Powered Agent Network for Task-Oriented Agent Collaboration	2023	arxiv	可作为动态组队和 agent 选择机制的早期 baseline。
Task-Oriented Collaboration & Dynamic Teaming	AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation	2023	arxiv	是 LLM 多 agent conversation framework 的基础论文，应作为协作运行时和对话协议 baseline。
Task-Oriented Collaboration & Dynamic Teaming	AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors in Agents	2023	ICLR 2024 Poster	补齐早期通用 LLM 多 agent 协作框架，对动态组队和正负涌现行为分析有参考价值。
Task-Oriented Collaboration & Dynamic Teaming	Talk Structurally, Act Hierarchically: A Collaborative Framework for LLM Multi-Agent Systems	2025	arxiv	直接补强“通信协议 + 分层协作”的设计路线，可作为动态团队协作的强 baseline。
Task-Oriented Collaboration & Dynamic Teaming	Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks	2024	arxiv	是 orchestrator-specialist 架构的强参考系统，也提供 AutoGenBench 这类可隔离评测工具。
Robustness & Reliability	On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents	2024	ICML 2025 Poster	直接对应项目“agent 上下线、能力波动、整体任务执行不稳定”的痛点。
Robustness & Reliability	VerifyMAS: Hypothesis Verification for Failure Attribution in LLM Multi-Agent Systems	2026	arxiv	自进化系统必须先知道失败来自分解、路由、执行、合成还是验证，故强相关。
Robustness & Reliability	Maris: A Formally Verifiable Privacy Policy Enforcement Paradigm for Multi-Agent Collaboration Systems	2025	arxiv	直接补强多 agent 通信的治理层，适合支撑敏感信息跨 agent 泄露和权限边界设计。
Task-Oriented Collaboration & Dynamic Teaming	Towards Effective GenAI Multi-Agent Collaboration: Design and Evaluation for Enterprise Applications	2024	arxiv	与项目落地到任务助手和企业级复杂任务高度贴合，尤其适合作端到端可用率与成本指标参考。
Task-Oriented Collaboration & Dynamic Teaming	Scaling Large Language Model-based Multi-Agent Collaboration	2024	ICLR 2025 Poster	是协作规模化和拓扑组织的重要论文，直接支撑“多个开源榜单 SOTA”和大规模协作框架目标。
Debate, Consensus & Verification	Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate	2023	EMNLP 2024	是 debate/verification 协作主线的基础论文，已有条目保留为核心协作机制。
Debate, Consensus & Verification	Improving Factuality and Reasoning in Language Models through Multiagent Debate	2023	arxiv	是 LLM multiagent debate 的代表性基础工作，应作为 debate 协作机制和 factuality 改进 baseline。
Debate, Consensus & Verification	ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs	2023	ACL 2024	对异构 agent 的 consensus 机制、置信度字段和投票聚合策略很有参考价值。
Debate, Consensus & Verification	Mixture-of-Agents Enhances Large Language Model Capabilities	2024	ICLR 2025 Spotlight	是多模型/多 agent 层级聚合的关键 baseline，也暴露了后续稀疏路由和成本控制需求。
Debate, Consensus & Verification	GroupDebate: Enhancing the Efficiency of Multi-Agent Debate Using Group Discussion	2024	arxiv	直接对应项目“协作不能增加 token/耗时”的成本约束。
Debate, Consensus & Verification	Stay Focused: Problem Drift in Multi-Agent Debate	2025	EACL 2025	重要负面结果，提醒协作轮次越多不一定越好，必须有终止和聚焦机制。
Debate, Consensus & Verification	iMAD: Intelligent Multi-Agent Debate for Efficient and Accurate LLM Inference	2025	AAAI 2026	非常贴合“按需协作”和“成本不增”的产品目标。
Debate, Consensus & Verification	ARMOR-MAD: Adaptive Routing for Heterogeneous Multi-Agent Debate in Large Language Model Reasoning	2026	arxiv	与项目的“能力波动、动态路由、成本控制”高度相关。
Software Engineering Multi-Agent Collaboration	ChatDev: Communicative Agents for Software Development	2023	ACL 2024	是软件工程多 agent 协作的代表性工作，应作为固定流程和语言通信协作 baseline。
Software Engineering Multi-Agent Collaboration	Co-Saving: Resource Aware Multi-Agent Collaboration for Software Development	2025	arxiv	直接服务项目“成本不增”和“从历史轨迹进化”的目标。
Software Engineering Multi-Agent Collaboration	Optimizing LLM-Based Multi-Agent System with Textual Feedback: A Case Study on Software Development	2025	arxiv	是“根据失败解释优化多 agent 角色/prompt”的直接相关论文。
Software Engineering Multi-Agent Collaboration	Self-Organizing Multi-Agent Systems for Continuous Software Development	2026	arxiv	强相关于动态 agent 组队、长期任务和自组织协作。
Software Engineering Multi-Agent Collaboration	Towards Engineering Multi-Agent LLMs: A Protocol-Driven Approach	2025	APSEC 2025	与“通用协作框架”和 A2A/MCP 风格协议化设计高度贴合。
Evolution & Optimization	Cross-Task Experiential Learning on LLM-based Multi-Agent Collaboration	2025	arxiv	直接补强“跨任务经验积累”和“协作系统进化”方向。
Evolution & Optimization	GPTSwarm: Language Agents as Optimizable Graphs	2024	ICML 2024 Oral	几乎直接对应“多 agent 反馈链路如何优化”，可作为群体进化主 baseline。
Evolution & Optimization	AFlow: Automating Agentic Workflow Generation	2024	ICLR 2025 Oral	适合作离线 workflow 自动优化器，对项目“可自进化协作系统”非常关键。
Evolution & Optimization	EvoAgent: Towards Automatic Multi-Agent Generation via Evolutionary Algorithms	2024	NAACL 2025	与项目“群体进化”命名和目标直接对齐。
Evolution & Optimization	Multi-Agent Collaboration via Evolving Orchestration	2025	NeurIPS 2025	直接补齐“协作编排本身可进化”的顶会工作，适合和 GPTSwarm、AFlow、EvoAgent、EvoMAS 共同作为进化主线 baseline。
Evolution & Optimization	Evolutionary Generation of Multi-Agent Systems	2026	arxiv	对工程落地尤其重要，因为配置空间进化比任意代码进化更可控。
Evolution & Optimization	Evolve as a Team: Collaborative Self-Evolution for LLM-based Multi-Agent Systems	2026	arxiv	最贴合“多 agent 之间互相影响的复杂反馈链路”的研究问题。
Evolution & Optimization	Swarm Skills: A Portable, Self-Evolving Multi-Agent System Specification	2026	arxiv	对“通用协作框架、可迁移到其他智能体产品线”非常贴合。
Evaluation & Benchmarks	Agent-as-a-Judge: Evaluate Agents with Agents	2024	ICML 2025 Poster	自进化需要 reward/critic，该论文可支撑反馈信号设计。
Communication & Routing	Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems	2025	ICLR 2025	直接命中“协作不能显著增加 token/成本”的工程边界，是通信剪枝和经济型协作拓扑的关键顶会论文。
Communication & Routing	G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks	2025	ICML 2025	与动态通信网络、任务自适应路由和低成本协作高度相关，可与 GPTSwarm、AgentPrune、MasRouter 形成拓扑优化主线。
Communication & Routing	MasRouter: Learning to Route LLMs for Multi-Agent Systems	2025	ACL 2025	直接服务“谁该参与、何时参与、以什么成本参与”的路由问题，是多 agent 通信与动态组队的关键 ACL 论文。
Task-Oriented Collaboration & Dynamic Teaming	Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation	2025	AAAI 2026 Oral	直接补强自动组队与通信拓扑生成，是从固定团队走向按任务生成团队的高质量前沿工作。
Evolution & Optimization	MAS-GPT: Training LLMs to Build LLM-based Multi-Agent Systems	2025	ICML 2025 Poster	直接面向“根据任务自动生成协作系统”的进化方向，可与 AFlow、SwarmAgentic、ARG-Designer 共同构成自动化设计主线。
Evaluation & Benchmarks	MultiAgentBench: Evaluating the Collaboration and Competition of LLM agents	2025	ACL 2025	是当前多 agent 协作评测的重要 ACL 基准，可用于验证通信拓扑和协作策略是否真正提升任务完成与协作质量。

Core GitHub Projects and Systems

Area	Project	Why it matters
Frameworks & Runtimes	LangGraph	适合做多 agent 协作运行时和 traceable workflow baseline。
Frameworks & Runtimes	AutoGen	适合参考多 agent conversation、事件驱动通信和分布式 agent runtime。
Frameworks & Runtimes	AgentScope	与“通用协作框架 + 动态协作 + 鲁棒部署”高度贴合。
Frameworks & Runtimes	CAMEL	适合做多 agent 实验平台和 agent society baseline。
Frameworks & Runtimes	ChatDev	是软件工程多 agent 协作和协作 scaling law 的核心参考实现。
Protocols & Interoperability	Model Context Protocol	多智能体框架若要跨产品线复用，工具/上下文接入层建议兼容 MCP。
Protocols & Interoperability	Agent2Agent Protocol	与“通用协作框架”和跨产品线 agent 协作高度相关。
Benchmarks & Evaluation	Multi-Agents-Debate	是 debate 协作机制的核心 baseline。
Research Implementations & Baselines	TalkHier	适合做“结构化消息协议 + 分层协作修正”的核心 baseline。
Research Implementations & Baselines	ReConcile	可作为异构 agent 共识、置信度表达和投票聚合的基础实现。
Research Implementations & Baselines	GPTSwarm	是群体进化和“多 agent 反馈链路优化”的核心实现，应作为主 baseline。
Reference Systems & Case Studies	Magentic-One	是 orchestrator-subagent 架构、错误恢复和隔离评测的高价值参考系统。
Research Implementations & Baselines	AgentPrune	是低成本通信和通信鲁棒性实验的重要 baseline，适合直接纳入复现实验。
Research Implementations & Baselines	MasRouter	是动态路由、低成本协作和模型选择实验的核心 baseline。
Benchmarks & Evaluation	MARBLE / MultiAgentBench	对验证多 agent 通信拓扑、协作协议和竞争/协作行为质量非常关键。
Research Implementations & Baselines	MAS-GPT	是自动化多 agent 系统生成和进化方向的重要实现。
Research Implementations & Baselines	ARG-Designer	对自动组队、自动拓扑设计和低成本协作非常关键，适合作为前沿 baseline。

Core Technical Readings

Group	Reading	Why it matters
Anthropic Official Engineering Articles	Anthropic - How we built our multi-agent research system	是“多步骤信息搜集+整合”场景最直接的公开工程案例，应作为系统设计和成本评估的必读材料。
Anthropic Official Engineering Articles	Anthropic - Building Effective AI Agents	是多 agent 协作模式选择的基础阅读，能防止把所有任务都错误地做成复杂多 agent。
Anthropic Official Engineering Articles	Anthropic - Building multi-agent systems	直接约束本项目“token 消耗和端到端耗时维持不增”的工程边界。
Anthropic Official Engineering Articles	Anthropic - Effective context engineering for AI agents	多 agent 通信路由本质上依赖上下文预算和信息选择，这篇可直接支撑通信层设计。
Anthropic Official Engineering Articles	Anthropic - Multi-agent coordination patterns	可作为本仓库协作分类和系统架构章节的工程模式骨架。
OpenAI Official Engineering Articles and Docs	OpenAI - Harness engineering	为多 agent 工程化提供 harness、验证和可靠性约束，是系统落地的重要方法论。
OpenAI Official Engineering Articles and Docs	OpenAI Agents SDK - Orchestrating Multiple Agents	对用 OpenAI Agents SDK 复现 supervisor/manager/handoff 协作模式非常关键。

Core Non-GitHub References

Area	Reference	Why it matters
Protocols & Interoperability	SEMAP	虽主要是论文方案，但非常适合作为多 agent 协作协议设计参考。

Complete Paper Catalog

Communication & Routing

Paper	Year	Venue / source	Status	Tags	Why it matters
Learning Multiagent Communication with Backpropagation	2016	NeurIPS 2016	foundational	commnet, differentiable-communication, marl	是“通信可以被学习”的基础论文，可作为 LLM 多 agent 路由学习的理论背景。
Learning to Communicate with Deep Multi-Agent Reinforcement Learning	2016	NeurIPS 2016	foundational	rial, dial, communication-learning, marl	支撑“通信协议可通过环境反馈优化”的理论基础。
Learning Attentional Communication for Multi-Agent Cooperation	2018	NeurIPS 2018	important	attentional-communication, marl, message-routing, cooperation	是“通信不是全连接广播，而是按需要形成通信组”的早期顶会基础论文。
Learning when to Communicate at Scale in Multiagent Cooperative and Competitive Tasks	2018	ICLR 2019	foundational	ic3net, communication-gating, scalable-marl	可借鉴为 LLM agent 通信开关、预算门控和 token 节省机制。
TarMAC: Targeted Multi-Agent Communication	2018	ICML 2019	foundational	targeted-communication, attention, marl	与项目中的“高效信息路由机制”高度同构。
Efficient Multi-Agent Communication via Shapley Message Value	2022	IJCAI 2022	important	message-value, shapley-value, communication-efficiency, marl	为“哪些消息值得发”提供可解释价值估计思路，适合连接 token budget 和通信收益评估。
Exchange-of-Thought: Enhancing Large Language Model Capabilities through Cross-Model Communication	2023	arxiv	important	cross-model-communication, relay, debate, memory	适合整理多 agent 通信协议设计空间。
A Scalable Communication Protocol for Networks of Large Language Models	2024	arxiv	supporting	communication-protocol, llm-networks, scalability, message-passing	为 A2A/ANP/ACP 等工程协议之外的学术协议设计提供参考，适合补齐协议层文献。
Improving Multi-Agent Debate with Sparse Communication Topology	2024	arxiv	core	sparse-topology, debate, communication-cost	支撑“不是所有 agent 都应该互相通信”的设计假设，适合做全连接通信的对照实验。
Learning Multi-Agent Communication from Graph Modeling Perspective	2024	ICLR 2024 Poster	important	learnable-graph, communication-topology, marl, graph-modeling	直接支撑通信拓扑可学习这一方向，可与 LLM 动态通信拓扑论文形成传统 MARL 到 LLM-MAS 的脉络。
SMoA: Improving Multi-agent Large Language Models with Sparse Mixture-of-Agents	2024	PAKDD 2025	important	sparse-mixture-of-agents, response-selection, early-stopping, role-diversity	是 MoA 类系统从全连接通信走向稀疏通信的直接参考，适合做成本受限多 agent 推理 baseline。
Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems	2025	ICLR 2025	core	agentprune, communication-pruning, token-efficiency, adversarial-robustness	直接命中“协作不能显著增加 token/成本”的工程边界，是通信剪枝和经济型协作拓扑的关键顶会论文。
Dynamic Generation of Multi-LLM Agents Communication Topologies with Graph Diffusion Models	2025	arxiv	core	topology-generation, graph-diffusion, cost-quality-robustness	适合支撑“动态 agent 组队 + 动态通信网络”的算法路线。
G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks	2025	ICML 2025	core	topology-design, graph-neural-network, task-adaptive-routing, token-efficiency	与动态通信网络、任务自适应路由和低成本协作高度相关，可与 GPTSwarm、AgentPrune、MasRouter 形成拓扑优化主线。
MasRouter: Learning to Route LLMs for Multi-Agent Systems	2025	ACL 2025	core	routing, llm-selection, multi-agent-systems, inference-efficiency	直接服务“谁该参与、何时参与、以什么成本参与”的路由问题，是多 agent 通信与动态组队的关键 ACL 论文。
RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory	2025	arxiv	core	context-routing, role-aware, structured-memory, token-budget	直接命中项目“通信代价大、关键信息传不准”的问题，可作为预算约束上下文路由主 baseline。
Differentiable Mixture-of-Agents Incentivizes Swarm Intelligence of Large Language Models	2026	arxiv	important	differentiable-routing, sparse-activation, test-time-adaptation, swarm-intelligence	连接“动态通信拓扑”和“在线自适应”，适合作为可学习路由和群体智能方向的前沿参考。
RADAR: Redundancy-Aware Diffusion for Multi-Agent Communication Structure Generation	2026	arxiv	important	redundancy-aware, diffusion, communication-structure-generation, topology-generation	补强 G-Designer、ARG-Designer 之后的通信结构生成路线，适合跟踪 2026 年前沿拓扑生成方法。
RouteMoA: Dynamic Routing without Pre-Inference Boosts Efficient Mixture-of-Agents	2026	arxiv	important	mixture-of-agents, dynamic-routing, model-selection, cost-latency	对“协作不能显著增加 token 和耗时”的约束非常关键，可支撑预推理前的 agent/model 路由。
TCAndon-Router: Adaptive Reasoning Router for Multi-Agent Collaboration	2026	arxiv	supporting	adaptive-routing, reasoning-router, collaboration, dynamic-selection	与 MasRouter、RCR-Router、ARMOR-MAD 共同构成动态路由方向，可作为后续候选 baseline。
The Five Ws of Multi-Agent Communication: Who Talks to Whom, When, What, and Why -- A Survey from MARL to Emergent Language and LLMs	2026	TMLR 2026	survey	survey, communication, marl, emergent-language, llm-agents	是本仓库通信章节的高质量分类法补充，可帮助把传统可学习通信和 LLM 自然语言通信放到同一脉络。
Towards Adaptive, Scalable, and Robust Coordination of LLM Agents: A Dynamic Ad-Hoc Networking Perspective	2026	arxiv	core	publish-subscribe, reputation, dynamic-networking, intent-routing	同时覆盖通信路由、动态协作和鲁棒性，是项目基础设施设计的强相关论文。

Task-Oriented Collaboration & Dynamic Teaming

Paper	Year	Venue / source	Status	Tags	Why it matters
Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments	2017	NeurIPS 2017	important	marl, actor-critic, mixed-cooperative-competitive, collaboration	是多智能体协作/竞争学习的重要基础论文，可作为 LLM agent 团队协作前的 MARL 基线背景。
Counterfactual Multi-Agent Policy Gradients	2018	AAAI 2018	important	marl, credit-assignment, counterfactual-baseline, cooperation	为多 agent 协作中的责任归因、失败归因和团队级优化提供基础算法背景。
QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning	2018	ICML 2018	important	marl, value-factorization, team-reward, cooperation	补齐协作任务中 team reward 和 credit assignment 的基础方法，有助于定义多 agent 协作评测指标。
Actor-Attention-Critic for Multi-Agent Reinforcement Learning	2019	ICML 2019	important	marl, attention, actor-critic, collaboration	与 LLM 多 agent 中“只关注相关 agent 输出”的思想一致，可作为协作注意力机制的基础参考。
A Dynamic LLM-Powered Agent Network for Task-Oriented Agent Collaboration	2023	arxiv	core	dynamic-agent-network, agent-selection, collaboration	可作为动态组队和 agent 选择机制的早期 baseline。
AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors in Agents	2023	ICLR 2024 Poster	core	agentverse, dynamic-composition, emergent-behavior, collaboration	补齐早期通用 LLM 多 agent 协作框架，对动态组队和正负涌现行为分析有参考价值。
AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation	2023	arxiv	core	autogen, multi-agent-conversation, tool-use, human-in-the-loop	是 LLM 多 agent conversation framework 的基础论文，应作为协作运行时和对话协议 baseline。
Theory of Mind for Multi-Agent Collaboration via Large Language Models	2023	EMNLP 2023	important	theory-of-mind, belief-state, cooperative-game, collaboration	适合支撑“agent 之间需要理解彼此状态和意图”的协作机制章节。
Adaptive In-conversation Team Building for Language Model Agents	2024	arxiv	important	captainagent, team-building, agent-selection, in-conversation	补齐动态组队中“对话中途扩队/换队”的机制，可作为 AgentVerse、Dynamic Agent Network 之后的关键候选。
AgentCoord: Visually Exploring Coordination Strategy for LLM-based Multi-Agent Collaboration	2024	arxiv	important	coordination-strategy, visual-exploration, structured-representation, human-in-the-loop	适合支撑“协作协议/拓扑如何设计和调试”的工具化方向。
Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks	2024	arxiv	core	orchestrator, specialist-agents, replanning, autogenbench	是 orchestrator-specialist 架构的强参考系统，也提供 AutoGenBench 这类可隔离评测工具。
MedAide: Information Fusion and Anatomy of Medical Intents via LLM-based Agent Collaboration	2024	Information Fusion 2025	supporting	medical, intent-aware, role-rotation, information-fusion	应用较窄，但 intent-aware fusion 和 role rotation 对通用协作框架有借鉴价值。
Multi-Agent Collaboration in Incident Response with Large Language Models	2024	arxiv	important	incident-response, cybersecurity, centralized, decentralized, hybrid	可作为动态协作结构和高压任务响应的应用案例。
Scaling Large Language Model-based Multi-Agent Collaboration	2024	ICLR 2025 Poster	core	macnet, scaling-law, dag-topology, thousand-agents	是协作规模化和拓扑组织的重要论文，直接支撑“多个开源榜单 SOTA”和大规模协作框架目标。
Towards Effective GenAI Multi-Agent Collaboration: Design and Evaluation for Enterprise Applications	2024	arxiv	core	enterprise, coordination-mode, routing-mode, payload-reference	与项目落地到任务助手和企业级复杂任务高度贴合，尤其适合作端到端可用率与成本指标参考。
Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation	2025	AAAI 2026 Oral	core	arg-designer, autoregressive-graph-generation, topology-design, role-selection	直接补强自动组队与通信拓扑生成，是从固定团队走向按任务生成团队的高质量前沿工作。
Collaborating Action by Action: A Multi-agent LLM Framework for Embodied Reasoning	2025	arxiv	important	mindcraft, minecollab, embodied-agents, communication-efficiency	是“多 agent 通信并不总是越多越好”的重要负面证据，直接支撑通信压缩、意图路由和协作成本控制。
Collaborative Belief Reasoning with LLMs for Efficient Multi-Agent Collaboration	2025	arxiv	important	belief-modeling, intent-inference, embodied-collaboration, communication-efficiency	直接补强“协作者意图理解 + 低通信成本协作”，适合和 Theory of Mind、belief state 类工作放在同一协作脉络。
Mediator-Guided Multi-Agent Collaboration among Open-Source Models for Medical Decision-Making	2025	arxiv	supporting	mediator-agent, heterogeneous-models, medical-vqa, multimodal	适合借鉴 mediator-guided collaboration 和异构模型协作机制。
Talk Structurally, Act Hierarchically: A Collaborative Framework for LLM Multi-Agent Systems	2025	arxiv	core	structured-communication, hierarchical-refinement, talkhier, collaboration	直接补强“通信协议 + 分层协作”的设计路线，可作为动态团队协作的强 baseline。
Tiered Agentic Oversight: A Hierarchical Multi-Agent System for Healthcare Safety	2025	arxiv	important	hierarchical-oversight, safety, routing, healthcare	对项目的 hierarchical routing、error absorption、human-in-the-loop 协作有参考价值。
ConSensus: Efficient Multi-Agent Collaboration for Multimodal Sensing	2026	arxiv	important	multimodal, consensus, hybrid-fusion, token-efficient	是“专业 agent + 低成本融合”的好例子，适合任务助手多源信息整合场景。

Debate, Consensus & Verification

Paper	Year	Venue / source	Status	Tags	Why it matters
Can ChatGPT Defend its Belief in Truth? Evaluating LLM Reasoning via Debate	2023	Findings of EMNLP 2023	important	debate, reasoning-evaluation, belief-defense, verification	补充多 agent 辩论的评测视角，可用于分析共识形成是否真正提升事实性和推理。
ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate	2023	arxiv	important	multi-agent-evaluation, debate, llm-as-judge	可用于项目中的结果验证和 agent-as-judge 设计。
Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate	2023	EMNLP 2024	core	mad, divergent-thinking, judge, reasoning	是 debate/verification 协作主线的基础论文，已有条目保留为核心协作机制。
Improving Factuality and Reasoning in Language Models through Multiagent Debate	2023	arxiv	core	multiagent-debate, factuality, reasoning, society-of-minds	是 LLM multiagent debate 的代表性基础工作，应作为 debate 协作机制和 factuality 改进 baseline。
ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs	2023	ACL 2024	core	round-table, consensus, diverse-llms, confidence-weighted-voting	对异构 agent 的 consensus 机制、置信度字段和投票聚合策略很有参考价值。
GroupDebate: Enhancing the Efficiency of Multi-Agent Debate Using Group Discussion	2024	arxiv	core	group-debate, token-efficient, debate-groups	直接对应项目“协作不能增加 token/耗时”的成本约束。
Mixture-of-Agents Enhances Large Language Model Capabilities	2024	ICLR 2025 Spotlight	core	mixture-of-agents, layered-aggregation, multi-model-collaboration, ensemble	是多模型/多 agent 层级聚合的关键 baseline，也暴露了后续稀疏路由和成本控制需求。
More Agents Is All You Need	2024	TMLR 2024	important	agent-forest, sampling-and-voting, scaling-agents, ensemble	提供“更多 agent 是否值得”的基础 scaling 对照，可作为复杂协作机制的低成本 baseline。
Enhancing Multi-Agent Debate System Performance via Confidence Expression	2025	EMNLP 2025	important	confidence-expression, debate-dynamics, calibration	可用于设计 agent 消息协议中的 confidence 字段和动态仲裁机制。
iMAD: Intelligent Multi-Agent Debate for Efficient and Accurate LLM Inference	2025	AAAI 2026	core	selective-debate, trigger-policy, token-efficient, confidence	非常贴合“按需协作”和“成本不增”的产品目标。
Stay Focused: Problem Drift in Multi-Agent Debate	2025	EACL 2025	core	problem-drift, debate-failure, driftjudge, driftpolicy	重要负面结果，提醒协作轮次越多不一定越好，必须有终止和聚焦机制。
ARMOR-MAD: Adaptive Routing for Heterogeneous Multi-Agent Debate in Large Language Model Reasoning	2026	arxiv	core	adaptive-routing, heterogeneous-agents, debate	与项目的“能力波动、动态路由、成本控制”高度相关。
Courtroom-Style Multi-Agent Debate with Progressive RAG and Role-Switching for Controversial Claim Verification	2026	arxiv	important	courtroom-debate, progressive-rag, role-switching, claim-verification	对结构化 adversarial collaboration 和证据扩展机制有参考价值。
DynaDebate: Breaking Homogeneity in Multi-Agent Debate with Dynamic Path Generation	2026	arxiv	important	dynamic-path-generation, process-centric-debate, verification-agent	对“动态角色/路径分配 + verifier 触发机制”很有参考价值。
Multi-Agent Debate with Memory Masking	2026	arxiv	important	memory-masking, debate-robustness, erroneous-memory	直接支撑“关键信息传不准/错误信息传播”的上下文治理机制。
Tool-MAD: A Multi-Agent Debate Framework for Fact Verification with Diverse Tool Augmentation and Adaptive Retrieval	2026	arxiv	important	tool-augmented-debate, adaptive-retrieval, fact-verification	适合任务助手中的事实核验、工具专长协作和 evidence-aware debate。

Role-Playing, Agent Society & Simulation

Paper	Year	Venue / source	Status	Tags	Why it matters
CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society	2023	NeurIPS 2023	foundational	camel, role-playing, inception-prompting, agent-society	是 LLM 多 agent 角色协作和 agent society 的基础论文，应作为后续 CAMEL/role-playing 工作源头。
Generative Agents: Interactive Simulacra of Human Behavior	2023	UIST 2023	foundational	generative-agents, memory, reflection, social-simulation	对长期记忆、反思和群体涌现行为建模有基础价值，可支撑“协作系统如何积累经验”的背景。
SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents	2023	ICLR 2024 Spotlight	important	social-intelligence, role-play, interaction-evaluation, sotopia	可补足多 agent 协作评测中“社会智能、策略沟通、角色互动”这类非代码任务维度。
LLM Harmony: Multi-Agent Communication for Problem Solving	2024	arxiv	supporting	role-playing, personas, communication, problem-solving	可作为早期 role-playing collaboration baseline。

Software Engineering Multi-Agent Collaboration

Paper	Year	Venue / source	Status	Tags	Why it matters
ChatDev: Communicative Agents for Software Development	2023	ACL 2024	core	chatdev, chat-chain, communicative-dehallucination, software-development	是软件工程多 agent 协作的代表性工作，应作为固定流程和语言通信协作 baseline。
MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework	2023	arxiv	important	sop, role-specialization, software-agents	适合作固定 SOP 协作 baseline，用于对比动态拓扑和自进化方法。
AgentMesh: A Cooperative Multi-Agent Generative AI Framework for Software Development Automation	2025	arxiv	important	planner, coder, debugger, reviewer, software	是经典角色分工流水线 baseline，适合和动态/进化式协作比较。
Co-Saving: Resource Aware Multi-Agent Collaboration for Software Development	2025	arxiv	core	resource-aware, shortcuts, experiential-knowledge, token-reduction	直接服务项目“成本不增”和“从历史轨迹进化”的目标。
DocAgent: A Multi-Agent System for Automated Code Documentation Generation	2025	arxiv	important	documentation, topological-code-processing, verifier, orchestrator	适合参考“上下文拓扑排序 + 专业 agent + verifier”的协作结构。
Optimizing LLM-Based Multi-Agent System with Textual Feedback: A Case Study on Software Development	2025	arxiv	core	textual-feedback, prompt-optimization, group-optimization, software	是“根据失败解释优化多 agent 角色/prompt”的直接相关论文。
Shadows in the Code: Exploring the Risks and Defenses of LLM-based Multi-Agent Software Development Systems	2025	AAAI 2026	important	security, malicious-agent, software-development, defense	说明多 agent 协作系统必须有分角色安全和关键 agent 防护。
Towards Engineering Multi-Agent LLMs: A Protocol-Driven Approach	2025	APSEC 2025	core	protocol, structured-messaging, lifecycle, a2a, semap	与“通用协作框架”和 A2A/MCP 风格协议化设计高度贴合。
Self-Organizing Multi-Agent Systems for Continuous Software Development	2026	arxiv	core	self-organizing, continuous-development, manager-agents, hire-fire	强相关于动态 agent 组队、长期任务和自组织协作。

Robustness & Reliability

Paper	Year	Venue / source	Status	Tags	Why it matters
On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents	2024	ICML 2025 Poster	core	faulty-agents, resilience, challenge, reviewer	直接对应项目“agent 上下线、能力波动、整体任务执行不稳定”的痛点。
Amplified Vulnerabilities: Structured Jailbreak Attacks on LLM-based Multi-Agent Debate	2025	arxiv	important	jailbreak, multi-agent-debate, security, attack	说明多 agent 协作不天然更安全，必须把消息安全、角色安全和验证机制纳入设计。
Disagreements Can Help: A Self-Repair Approach to LLM Agents' Robustness in Knowledge Conflicts	2025	arxiv	important	knowledge-conflict, self-repair, disagreement	可用于设计冲突检测、互相质疑和自修复协作机制。
GUARDIAN: Safeguarding LLM Multi-Agent Collaborations with Temporal Graph Modeling	2025	arxiv	important	temporal-graph, safety, hallucination-propagation, anomaly-detection	直接服务于“关键信息传不准”和错误跨 agent 放大的安全问题，可作为 trace graph 防御方向。
Maris: A Formally Verifiable Privacy Policy Enforcement Paradigm for Multi-Agent Collaboration Systems	2025	arxiv	core	privacy-policy, message-flow-control, reference-monitor, prompt-injection	直接补强多 agent 通信的治理层，适合支撑敏感信息跨 agent 泄露和权限边界设计。
VerifyMAS: Hypothesis Verification for Failure Attribution in LLM Multi-Agent Systems	2026	arxiv	core	failure-attribution, trace-analysis, verification	自进化系统必须先知道失败来自分解、路由、执行、合成还是验证，故强相关。
When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems	2026	arxiv	important	safety, malicious-agent, message-defense	适合支撑多 agent 消息污染、防御和可信通信子方向。

Evolution & Optimization

Paper	Year	Venue / source	Status	Tags	Why it matters
Reflexion: Language Agents with Verbal Reinforcement Learning	2023	NeurIPS 2023	foundational	self-reflection, verbal-rl, episodic-memory	是自进化机制的单 agent 基础，可扩展到团队级经验沉淀。
Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation	2023	arxiv	foundational	self-improvement, scaffold-optimization, code-generation	可作为自动改 prompt/tool/workflow 的思想背景。
Voyager: An Open-Ended Embodied Agent with Large Language Models	2023	TMLR 2024	foundational	lifelong-learning, skill-library, self-verification	支撑“技能库/经验库”式长期进化，但本身不是多 agent 主线。
AFlow: Automating Agentic Workflow Generation	2024	ICLR 2025 Oral	core	workflow-search, mcts, agentic-workflow	适合作离线 workflow 自动优化器，对项目“可自进化协作系统”非常关键。
EvoAgent: Towards Automatic Multi-Agent Generation via Evolutionary Algorithms	2024	NAACL 2025	core	evolutionary-algorithm, multi-agent-generation, mutation, crossover	与项目“群体进化”命名和目标直接对齐。
ExpeL: LLM Agents Are Experiential Learners	2024	AAAI 2024	important	experiential-learning, memory, reflection, agent-learning	虽然不是多 agent 专用，但为“从执行轨迹中积累经验并改进行为”的进化闭环提供关键基线。
Experiential Co-Learning of Software-Developing Agents	2024	ACL 2024	important	experiential-learning, software-agents, co-learning, trajectory-memory	直接连接多 agent 软件协作和经验驱动进化，可作为 ChatDev 类系统如何从失败和轨迹中持续改进的核心参考。
GPTSwarm: Language Agents as Optimizable Graphs	2024	ICML 2024 Oral	core	gptswarm, graph-optimization, prompt-optimization, edge-optimization	几乎直接对应“多 agent 反馈链路如何优化”，可作为群体进化主 baseline。
AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems	2025	arxiv	important	decentralized-coordination, agent-evolution, rag, dynamic-routing	补齐去中心化协作进化路线，与 centralized orchestrator/manager 架构形成重要对照。
Cross-Task Experiential Learning on LLM-based Multi-Agent Collaboration	2025	arxiv	core	mael, cross-task-experience, graph-collaboration, reward-pool	直接补强“跨任务经验积累”和“协作系统进化”方向。
MAS-GPT: Training LLMs to Build LLM-based Multi-Agent Systems	2025	ICML 2025 Poster	core	mas-generation, query-adaptive, executable-code, system-design	直接面向“根据任务自动生成协作系统”的进化方向，可与 AFlow、SwarmAgentic、ARG-Designer 共同构成自动化设计主线。
Multi-Agent Collaboration via Evolving Orchestration	2025	NeurIPS 2025	core	evolving-orchestration, puppeteer, reinforcement-learning, dynamic-sequencing	直接补齐“协作编排本身可进化”的顶会工作，适合和 GPTSwarm、AFlow、EvoAgent、EvoMAS 共同作为进化主线 baseline。
SwarmAgentic: Towards Fully Automated Agentic System Generation via Swarm Intelligence	2025	EMNLP 2025	important	agentic-system-generation, swarm-intelligence, automated-design, evolution	直接补充多 agent 系统自动生成与进化方向，适合与 AFlow/GPTSwarm/EvoAgent 放在同一脉络。
Evolutionary Generation of Multi-Agent Systems	2026	arxiv	core	evomas, configuration-space, evolutionary-generation	对工程落地尤其重要，因为配置空间进化比任意代码进化更可控。
Evolve as a Team: Collaborative Self-Evolution for LLM-based Multi-Agent Systems	2026	arxiv	core	team-evolution, collaborative-self-evolution, post-task-communication	最贴合“多 agent 之间互相影响的复杂反馈链路”的研究问题。
Swarm Skills: A Portable, Self-Evolving Multi-Agent System Specification	2026	arxiv	core	swarm-skills, portable-specification, self-evolving	对“通用协作框架、可迁移到其他智能体产品线”非常贴合。

Evaluation & Benchmarks

Paper	Year	Venue / source	Status	Tags	Why it matters
AgentBench: Evaluating LLMs as Agents	2023	ICLR 2024 Poster	important	agent-benchmark, interactive-environment, evaluation	可作为任务助手多 agent 系统的公共能力评测参考。
GAIA: a benchmark for General AI Assistants	2023	ICLR 2024 Poster	important	general-assistant, web, tool-use, benchmark	适合作“多步骤信息搜集+整合”公共评测集。
SWE-bench: Can Language Models Resolve Real-World GitHub Issues?	2023	ICLR 2024 Oral	important	software-engineering, long-horizon, benchmark	适合测长上下文、多步骤协作和可恢复执行，但需补充非代码长尾任务。
Agent-as-a-Judge: Evaluate Agents with Agents	2024	ICML 2025 Poster	core	agent-evaluation, process-feedback, llm-judge	自进化需要 reward/critic，该论文可支撑反馈信号设计。
VillagerAgent: A Graph-Based Multi-Agent Framework for Coordinating Complex Task Dependencies in Minecraft	2024	arxiv	important	villagerbench, minecraft, dag-coordination, complex-dependencies	适合作为复杂依赖、多 agent 分工、同步和动态适应能力的协作评测场景。
MALLM: Multi-Agent Large Language Models Framework	2025	EMNLP 2025 Demo	important	multi-agent-debate, framework, configuration, evaluation-pipeline	适合系统比较 debate、relay、memory、voting、consensus 等多 agent 策略，补齐“实验框架/可复现评测”维度。
MultiAgentBench: Evaluating the Collaboration and Competition of LLM agents	2025	ACL 2025	core	benchmark, collaboration, competition, coordination-protocols	是当前多 agent 协作评测的重要 ACL 基准，可用于验证通信拓扑和协作策略是否真正提升任务完成与协作质量。
MARBLE: Multi-Agent Reasoning for Bioinformatics Learning and Evolution	2026	arxiv	supporting	bioinformatics, multi-agent-reasoning, learning, evolution	虽是垂直领域，但对“领域复杂任务 + 多 agent 学习/进化”有参考价值，适合作补充性 benchmark。

Surveys & Taxonomies

Paper	Year	Venue / source	Status	Tags	Why it matters
A Survey of Multi-Agent Deep Reinforcement Learning with Communication	2024	Autonomous Agents and Multi-Agent Systems 2024	important	survey, marl, communication, protocols	为 LLM 多 agent 通信研究提供前 LLM 时代的系统背景，帮助区分学到的通信协议与提示式通信。
The Landscape of Emerging AI Agent Architectures for Reasoning, Planning, and Tool Calling	2024	arxiv	survey	survey, agent-architecture, planning, tool-use	适合作 agent architecture 背景章节。
Multi-Agent Collaboration Mechanisms: A Survey of LLMs	2025	arxiv	survey	survey, collaboration, coordination-protocols	适合作本综述的分类法参考。
Multi-Agent Coordination across Diverse Applications: A Survey	2025	arxiv	survey	survey, coordination, scalability, heterogeneity	适合支撑“协作”的理论定义和跨领域背景。

GitHub Project Catalog

Frameworks & Runtimes

Project	Status	Tags	Why it matters
AG2	important	autogen-fork, agentos, groupchat, multi-agent	适合与 Microsoft AutoGen 对比，观察社区版多 agent 框架如何演进。
Agent Squad	important	routing, multi-agent, specialist-agents, context	与项目“任务助手多专业 agent 联合执行”和 request routing 高度相关。
AgentScope	core	multi-agent-platform, message-exchange, fault-tolerance, distributed	与“通用协作框架 + 动态协作 + 鲁棒部署”高度贴合。
AgentVerse	important	multi-agent, framework, simulation, collaboration	可作为早期通用多 agent 协作框架和仿真实验平台参考。
Agno	important	agent-platform, teams, memory, runtime	适合参考多 agent team 在生产平台中的运行、管理和观测抽象。
AutoGen	core	multi-agent, event-driven, agentchat, runtime	适合参考多 agent conversation、事件驱动通信和分布式 agent runtime。
BeeAI Framework	important	python, typescript, workflows, production-agents	可作为企业级多 agent 运行时、工具接入和协议生态的补充参考。
CAMEL	core	agent-society, multi-agent, memory, benchmarks	适合做多 agent 实验平台和 agent society baseline。
ChatDev	core	software-company, chat-chain, multi-agent, macnet	是软件工程多 agent 协作和协作 scaling law 的核心参考实现。
CrewAI	important	crews, flows, role-based, business-workflows	可作为业务流程协作 baseline，但研究创新不应停留在框架调用。
deepagents	important	long-running, subagents, planning, langgraph	适合参考长尾复杂任务中 subagent 编排、任务持久化和工作状态管理。
Google ADK	important	agent-development-kit, multi-agent, evaluation, deployment	可作为多 agent 产品化 SDK、评测和部署路径的强参考。
LangGraph	core	graph-runtime, stateful, durable-execution, multi-agent	适合做多 agent 协作运行时和 traceable workflow baseline。
LlamaIndex	important	agent-workflows, retrieval, tools, multi-agent	对“多 agent + 检索/文档/数据任务”的任务助手落地很重要，适合作为 RAG-heavy 协作基线。
Mastra	important	typescript, workflows, agents, memory	对前端/Node 生态中多 agent workflow 与产品化部署很有参考价值。
MetaGPT	important	sop, role-specialization, software-company	适合做固定角色/固定流程协作 baseline。
Microsoft Agent Framework	important	multi-agent, workflows, orchestration, observability	适合跟踪 AutoGen 之后的生产级多 agent 编排方向。
NeMo Agent Toolkit	important	agent-toolkit, multi-agent, optimization, nvidia	与“多 agent 协作优化”和企业级 agent team 运行效率直接相关。
Open Multi-Agent	important	dag, tracing, typescript, multi-agent	适合作任务 DAG、并行协作和可观测执行 baseline。
OpenAI Agents SDK	important	agents, handoffs, tracing, guardrails, mcp	适合借鉴 handoff、manager-style orchestration 和 tracing 的简洁抽象。
OpenAI Swarm	important	educational, handoffs, routines, multi-agent-orchestration	虽非生产框架，但对理解 handoff、agent-as-tool 和轻量编排模式非常有参考价值。
PraisonAI	important	multi-agent, workforce, self-improving, rag	可作为业务流程中“多角色 agent workforce”落地方式的对照项目。
Pydantic AI	important	typed-agents, structured-output, graph, production	多 agent 协作需要强 schema、可验证状态和清晰工具契约，Pydantic AI 是重要工程参考。
Semantic Kernel	important	enterprise, orchestration, plugins, multi-agent	适合作企业场景中多 agent 编排、工具接入和结构化插件契约的工程参考。
smolagents	important	code-agents, managed-agents, tools, lightweight	对“简洁 agent loop + managed subagents”的最小实现有参考价值，可作为复杂框架的低开销对照。
Strands Agents	important	sdk, multi-agent, mcp, production	对“通用协作框架 + 跨模型/跨工具生态”有参考价值。

Protocols & Interoperability

Project	Status	Tags	Why it matters
Agent Client Protocol	important	acp, editor-agent, protocol, sessions	对跨宿主 agent 协作、agent handoff 和可迁移会话协议有参考价值。
Agent Communication Protocol	supporting	acp, agent-protocol, interoperability, archived	与 A2A、ANP、MCP 一起构成跨 agent 通信协议调研边界，应在协议对比中说明其归档状态。
Agent Network Protocol	important	anp, agent-network, interoperability, decentralized	补齐 A2A/MCP 之外的协议路线，适合研究跨组织 agent 发现和协作网络。
Agent2Agent Protocol	core	a2a, agent-interop, task-delegation, agent-cards	与“通用协作框架”和跨产品线 agent 协作高度相关。
Agora Protocol	supporting	agora, protocol, python-sdk, collaboration	可作为 ProtocolBench 类协议评测中的协议生态补充，不作为核心工程框架。
mcp-agent	important	mcp, workflow, multi-server, agent-framework	适合参考如何把多个工具服务器聚合进 agent 协作流程。
Model Context Protocol	core	mcp, tools, resources, prompts, interoperability	多智能体框架若要跨产品线复用，工具/上下文接入层建议兼容 MCP。

Benchmarks & Evaluation

Project	Status	Tags	Why it matters
AgentBench	important	agent-benchmark, environments, reasoning, decision-making	适合测基础 agent 执行能力，但需补充多 agent 通信和进化指标。
ChatEval	important	multi-agent-evaluator, debate, llm-as-judge	可作为多 agent verifier/judge 组件和评价方法 baseline。
Collaborative Gym	important	collaborative-agents, human-agent-collaboration, evaluation, environments	补齐 human-agent/team collaboration 评测维度，有助于评估多 agent 与人类协同的真实可用性。
DeepEval	supporting	evaluation, llm-judge, testing	可作为自建 Qianwen-LongTail-MAS-Bench 的评测基础设施参考。
E2EDev	important	end-to-end-software-development, bdd, multi-agent-annotation, benchmark	适合评估多 agent 从需求到交付的完整长任务能力。
M-MAD	supporting	multi-agent-debate, machine-translation-evaluation, llm-as-judge	适合参考如何把复杂评价标准拆成多个 specialist judges。
MARBLE / MultiAgentBench	core	multiagentbench, marble, acl-2025, collaboration, competition	对验证多 agent 通信拓扑、协作协议和竞争/协作行为质量非常关键。
Multi-Agents-Debate	core	multi-agent-debate, reasoning, judge	是 debate 协作机制的核心 baseline。
Promptfoo	supporting	eval, red-team, ci, regression	适合做多 agent 变更后的回归测试和安全测试。
SWE-bench	important	software-engineering, long-horizon, benchmark	可测长任务执行、上下文管理、验证和恢复能力。
SWE-Dev	important	feature-development, software-engineering, benchmark, rl	适合补充 SWE-bench 对“新增功能/端到端开发”的覆盖。
tau2-bench	important	tool-agent-user, benchmark, multi-turn, real-world-domains	适合评估多 agent 任务助手在真实工具和用户反馈循环中的协作表现。
Terminal-Bench	important	terminal, benchmark, long-horizon, verification	可用于测试多 agent 在代码、命令行、文件系统任务中的长程协作和恢复能力。
TheAgentCompany	important	workplace, simulated-company, benchmark, long-horizon	与多 agent 任务助手和软件团队协作评测高度相关。
WebArena	important	web-agent, benchmark, realistic-environment, tool-use	可作为任务助手“多步骤信息搜集+执行”的公共环境基准。
WorkArena	important	enterprise, web-agent, service-workflows, benchmark	适合补充企业长尾任务、流程系统和真实业务操作场景的评测。

Observability & Operations

Project	Status	Tags	Why it matters
AgentOps	important	agent-monitoring, tracing, cost, benchmarks	多 agent 通信、成本和失败归因需要统一轨迹与成本观测，AgentOps 是强工程参考。
Arize Phoenix	supporting	observability, eval, tracing, llmops	可作为 trace + evaluation + failure analysis 工程参考。
GUARDIAN	important	temporal-graph, safety, anomaly-detection, multi-agent	适合把多 agent 消息 trace 转成图并做错误传播/异常检测。
Langfuse	supporting	observability, traces, cost, prompts	适合记录多 agent 消息、成本、失败轨迹和进化前后对比。
LangSmith	supporting	observability, tracing, eval, debugging	多 agent 通信/进化需要 trace-native evaluation，LangSmith 可作工程参考。
LangWatch	important	evaluations, simulations, agent-testing, observability	适合作多 agent 变更后的离线仿真、回归评测和线上质量监控。
Weave	supporting	tracing, evaluations, llmops, wandb	可用于多 agent 轨迹回放、成本/质量对比和进化前后评估。

Research Implementations & Baselines

Project	Status	Tags	Why it matters
AgentForest	important	sampling-and-voting, agent-scaling, ensemble-baseline	适合作为最简单但必须比较的多 agent scaling baseline。
AgentNet	important	decentralized, evolutionary-coordination, rag, dynamic-routing	与中心化 orchestrator 架构形成对照，适合研究去中心化协作和进化机制。
AgentPrune	core	communication-pruning, token-efficiency, iclr-2025, robustness	是低成本通信和通信鲁棒性实验的重要 baseline，适合直接纳入复现实验。
ARG-Designer	core	aaai-2026, topology-generation, role-selection, graph-generation	对自动组队、自动拓扑设计和低成本协作非常关键，适合作为前沿 baseline。
Dr. MAS	important	reinforcement-learning, co-training, heterogeneous-llms, multi-agent	对群体进化和多 agent 学习机制有参考价值，适合作为训练型 MAS baseline。
Generative Agents	important	agent-society, memory, reflection, simulation	适合作为长期记忆、经验沉淀和 agent society 章节的经典参考实现。
GPTSwarm	core	optimizable-graphs, prompt-optimization, edge-optimization, evolution	是群体进化和“多 agent 反馈链路优化”的核心实现，应作为主 baseline。
MALLM	important	multi-agent-llms, consensus, debate, framework	适合快速搭建 debate、投票、共识类 baseline，并记录多 agent 实验配置。
MARTI	important	reinforced-training, inference, multi-agent-llm, optimization	补齐“多 agent 不只推理编排，也可训练/强化优化”的工程路线。
MAS-GPT	core	mas-generation, icml-2025, executable-code, auto-design	是自动化多 agent 系统生成和进化方向的重要实现。
MasRouter	core	routing, acl-2025, llm-selection, multi-agent	是动态路由、低成本协作和模型选择实验的核心 baseline。
RADAR	supporting	communication-structure, diffusion, redundancy-aware, topology-generation	作为 2026 通信结构生成候选 baseline，可与 G-Designer、ARG-Designer 对比。
ReConcile	core	round-table, consensus, confidence-weighted-voting, diverse-models	可作为异构 agent 共识、置信度表达和投票聚合的基础实现。
RMoA	important	residual-mixture-of-agents, diversity-selection, adaptive-termination	可作为 MoA 家族中“信息保真 + 成本控制”的重要对照。
SMoA	important	sparse-mixture-of-agents, response-selection, early-stopping	适合研究多 agent 聚合如何在保持效果的同时降低成本。
SOTOPIA	important	social-intelligence, role-play, interaction-evaluation	可补充非代码长尾任务中的协作、沟通和社会目标评测。
TalkHier	core	structured-communication, hierarchical-refinement, llm-multi-agent	适合做“结构化消息协议 + 分层协作修正”的核心 baseline。
VillagerAgent	important	minecraft, villagerbench, dag-coordination, complex-dependencies	适合构造任务依赖、同步、动态适应和 hallucination 控制的协作评测。

Reference Systems & Case Studies

Project	Status	Tags	Why it matters
DocAgent	important	documentation, verifier, orchestrator, code-understanding	是“专业 agent + topological context + verifier”的清晰参考系统。
Magentic-One	core	orchestrator, specialist-agents, autogenbench, replanning	是 orchestrator-subagent 架构、错误恢复和隔离评测的高价值参考系统。
OpenHands	important	software-agent, coding, benchmark, runtime	虽不是纯多 agent 框架，但可作为长任务执行、工具使用和软件任务评测的重要 baseline。
SWE-agent	important	software-agent, swe-bench, execution, repair	可作为单/多 agent 软件任务对比基线，帮助衡量多 agent 协作是否真正带来收益。
Symphony	important	orchestration, issue-driven, isolated-runs, coding-agents	虽偏编码 agent，但对“任务分解、隔离执行、编排控制面和回收结果”非常有参考价值。

Technical Docs, Blogs, and Field Reports

Anthropic Official Engineering Articles

Reading	Status	Tags	Why it matters
Anthropic - Building Effective AI Agents	core	engineering-guide, workflows, agents, patterns	是多 agent 协作模式选择的基础阅读，能防止把所有任务都错误地做成复杂多 agent。
Anthropic - Building multi-agent systems	core	field-report, context-protection, specialization, token-cost	直接约束本项目“token 消耗和端到端耗时维持不增”的工程边界。
Anthropic - Claude Code auto mode	supporting	permissions, autonomy, classifier, coding-agents	对多 agent 中 human-in-the-loop、权限门控和自动化边界设计有启发。
Anthropic - Claude Code sandboxing	supporting	sandboxing, security, autonomy, coding-agents	多 agent 协作会放大工具调用和文件系统风险，沙箱边界是基础治理机制。
Anthropic - Code execution with MCP	important	mcp, code-execution, sandbox, tool-boundaries	多 agent 系统需要清晰工具边界和执行隔离，MCP 是关键互操作层。
Anthropic - Demystifying evals for AI agents	important	agent-evals, nondeterminism, trajectories, evaluation	自进化多 agent 必须有可靠 eval/reward，这篇补齐评测方法论。
Anthropic - Effective context engineering for AI agents	core	context-engineering, memory, working-state, token-budget	多 agent 通信路由本质上依赖上下文预算和信息选择，这篇可直接支撑通信层设计。
Anthropic - Effective harnesses for long-running agents	important	long-running-agents, state, resumability, harness	任务助手长尾复杂任务通常是长时任务，多 agent 协作运行时必须处理恢复和状态漂移。
Anthropic - Harness design for long-running application development	important	long-running, app-development, harness, reliability	可为软件工程多 agent 协作和持续开发任务提供工程约束。
Anthropic - How we built our multi-agent research system	core	field-report, lead-agent, parallel-subagents, research	是“多步骤信息搜集+整合”场景最直接的公开工程案例，应作为系统设计和成本评估的必读材料。
Anthropic - Multi-agent coordination patterns	core	patterns, generator-verifier, orchestrator-subagent, message-bus, shared-state	可作为本仓库协作分类和系统架构章节的工程模式骨架。
Anthropic - Quantifying infrastructure noise in agentic coding evals	supporting	evaluation, infrastructure-noise, coding-agents, reproducibility	多 agent 系统评测更容易受环境和基础设施波动影响，这篇适合指导评测治理。
Anthropic - Scaling Managed Agents	important	managed-agents, session-logs, sandbox, architecture	对多 agent 运行时边界、隔离执行和状态回放设计很有参考价值。
Anthropic - Writing effective tools for AI agents	important	tool-design, schemas, agent-tools, reliability	多 agent 框架中的 specialist agents 常以工具形式互相调用，工具契约质量直接影响协作可靠性。

OpenAI Official Engineering Articles and Docs

Reading	Status	Tags	Why it matters
OpenAI - Building more helpful agents with a new evaluation framework	important	field-report, browsecomp, evaluation, hard-to-find-information	适合支撑“多步骤信息搜集+整合”类任务助手评测，尤其是高难检索和证据链质量。
OpenAI - Harness engineering	core	harness-engineering, verification, reliability, agent-first-software	为多 agent 工程化提供 harness、验证和可靠性约束，是系统落地的重要方法论。
OpenAI - Symphony orchestration	important	orchestration, codex, issue-driven, control-plane	可借鉴到多 agent 任务助手的任务队列、隔离执行和结果合并流程。
OpenAI - The next evolution of the Agents SDK	important	agents-sdk, sandbox, manifests, memory, tools	对生产级 agent SDK 如何标准化工具、记忆和执行环境有参考价值。
OpenAI - Unrolling the Codex agent loop	important	agent-loop, tool-replay, stateless-calls, codex	多 agent 系统同样需要处理轨迹回放、上下文增长和工具调用状态，这篇能支撑运行时设计。
OpenAI Agents SDK - Orchestrating Multiple Agents	core	documentation, agents-sdk, handoffs, orchestration	对用 OpenAI Agents SDK 复现 supervisor/manager/handoff 协作模式非常关键。
OpenAI Cookbook - Orchestrating Agents	important	cookbook, handoffs, routines, orchestration	是 OpenAI 生态中最直接的多 agent 编排实践示例，应和 Swarm/Agents SDK 一起参考。
OpenAI Developers - Testing Agent Skills Systematically with Evals	important	skills, evals, traces, regression	自进化多 agent 系统需要 regression-safe 的技能/协作模式评测闭环。

Multi-Agent Framework Documentation

Reading	Status	Tags	Why it matters
AG2 Documentation - Multi-agent Chat	supporting	documentation, ag2, groupchat, multi-agent	适合与 AutoGen 官方模式对照，观察 group chat 协作抽象的演进。
AutoGen Documentation - Multi-agent Design Patterns	important	documentation, autogen, design-patterns, runtime	适合把 AutoGen 论文和 Magentic-One 系统落到可复现实验配置。
Hugging Face smolagents - Multi-agent Systems	important	documentation, smolagents, managed-agents, code-agents	适合用最小 agent 框架验证“manager + specialist”的协作开销和效果。
LangGraph Documentation - Multi-agent Systems	important	documentation, langgraph, supervisor, swarm	是构建 traceable multi-agent workflow baseline 的直接工程参考。
LlamaIndex Documentation - Multi-agent Workflows	important	documentation, llamaindex, workflows, retrieval-agents	对多 agent 任务助手中的数据检索、工具调用和 workflow 组合有直接工程价值。
Mastra Documentation - Agent Networks	important	documentation, mastra, agent-network, typescript	补齐 Node/TypeScript 生态下多 agent network 工程实践。
Pydantic AI Documentation - Multi-agent Applications	important	documentation, pydantic-ai, typed-agents, applications	对要求高可靠 schema、工具参数和状态校验的多 agent 系统尤其有价值。

Protocol Documentation

Reading	Status	Tags	Why it matters
Google A2A Protocol Documentation	important	documentation, a2a, interoperability, agent-cards	对“通用协作框架”和跨产品线 agent 协作协议设计高度相关。
Model Context Protocol Documentation	important	documentation, mcp, tools, context	多 agent 系统若要可迁移和可扩展，工具/上下文层应优先兼容 MCP。
ProtocolBench - Which LLM Multi-Agent Protocol to Choose?	supporting	protocolbench, protocolrouter, a2a, anp, acp, mcp	虽未作为正式顶会论文收录，但对协议选型、协议评测指标和 hybrid protocol routing 很有参考价值。

Evaluation, Runtime, and Engineering Blogs

Reading	Status	Tags	Why it matters
Cognition - Don't Build Multi-Agents	important	anti-patterns, agent-design, reliability, single-agent-baseline	是重要负面/约束性阅读，可帮助判断什么时候不该上多 agent，防止复杂度和成本无谓膨胀。
Cognition - What We Learned Building Cloud Agents	important	cloud-agents, vm-isolation, snapshots, governance	对多 agent 长任务运行环境、隔离执行和企业落地非常有参考价值。
HumanLayer - 12 Factor Agents	important	production-agents, principles, operations, reliability	适合作多 agent 任务助手工程化 checklist。
HumanLayer - Skill Issue: Harness Engineering for Coding Agents	supporting	harness-engineering, coding-agents, skills, operations	可迁移到多 agent 中 skill/workflow/tool 描述的系统化管理。
Inngest - Your Agent Needs a Harness, Not a Framework	important	harness-first, reliability, orchestration, production	对多 agent 系统的队列、重试、状态和事件驱动 orchestration 很有启发。
LangChain - Agent frameworks, runtimes, and harnesses	important	architecture, framework, runtime, harness	有助于把多 agent 通信/协作算法和底层 harness 工程职责分离。
LangChain - Evaluating Deep Agents	important	deep-agents, evaluation, long-horizon, stateful-agents	适合设计多 agent 长尾任务评测、错误分类和回归测试。
LangChain - How and when to build multi-agent systems	important	engineering-guide, supervisor, handoff, multi-agent	可帮助把论文中的协作模式映射到 LangGraph/LangChain 工程实现。
LangChain - Improving Deep Agents with harness engineering	important	deep-agents, harness-engineering, benchmark, iteration	支撑“自进化不一定先改模型，可以先优化 harness/workflow”的研究路线。
LangChain - The Anatomy of an Agent Harness	important	harness, architecture, components, reliability	是把多 agent 框架落成可维护工程系统的架构参考。
Martin Fowler - Harness Engineering	important	architecture, harness-engineering, entropy, gen-ai	为多 agent 系统边界、治理和工程复杂度控制提供高质量架构视角。

High-Quality Personal and Community Blogs

Reading	Status	Tags	Why it matters
DeepLearning.AI - Four AI Agent Strategies	important	agentic-design-patterns, multi-agent-collaboration, reflection, planning	是工业界传播最广的 agentic pattern 框架之一，适合放在多 agent 协作模式章节作为概念入口。
LessWrong - Survey of Multi-agent LLM Evaluations	supporting	multi-agent-evaluation, risks, miscoordination, collusion	补齐多 agent 评测不只看任务分数，还要看协作失败、串谋和安全风险的视角。
Lilian Weng - LLM Powered Autonomous Agents	important	agent-overview, planning, memory, tool-use, reflection	虽不是多 agent 专文，但为理解多 agent 中每个 agent 的能力边界和自我改进机制提供基础。
Simon Willison - Agent Definition	supporting	agent-definition, tool-loop, terminology, multi-agent	有助于统一“agent”和“multi-agent system”的工程术语，避免把普通 workflow 误称为多 agent。

Non-GitHub Project and Benchmark References

Protocols & Interoperability

Reference	Status	Tags	Why it matters
SEMAP	core	protocol, structured-messaging, a2a, software-engineering	虽主要是论文方案，但非常适合作为多 agent 协作协议设计参考。

Benchmarks & Evaluation

Reference	Status	Tags	Why it matters
GAIA	important	general-assistant, web, tool-use, benchmark	适合任务助手信息搜集、整合和工具使用评测。

Observability & Operations

Reference	Status	Tags	Why it matters
AgentCoord	supporting	coordination-visualization, strategy-design, human-in-the-loop	适合作协作拓扑、任务依赖和 agent 关系的调试工具参考。

Research Implementations & Baselines

Reference	Status	Tags	Why it matters
MINDCraft / MineCollab	important	minecraft, communication-efficiency, embodied-agents, collaboration-benchmark	可作为 embodied multi-agent 场景下“通信越多是否越好”和“协作是否真正带来收益”的强评测补充。

Maintenance

Source of truth:

data/papers.yaml
data/projects.yaml

Regenerate generated files:

python3 scripts/render.py

Validate schema and links:

python3 scripts/verify.py

Verify GitHub-backed project entries with GitHub CLI:

python3 scripts/verify_github_projects.py

Similar repos

curation-labs/awesome-agent-coordination

kyegomez/awesome-multi-agent-papers

A compilation of the best multi-agent papers

1,689

vchaitanyachowdari/awesome-multi-agent-papers

A Compilation of the best Multi Agent Papers

luo-junyu/Awesome-Agent-Papers

[Up-to-date] Large Language Model Agent: A Survey on Methodology, Applications and Challenges

2,862

Yigtwxx/awesome-claude-multi-agent

A curated list of frameworks, patterns, and research for orchestrating multi-agent systems built on…

johnsutor/awesome-multi-agent-optimization

A collection of papers and libraries for performing multi-agent optimization

EvoMap/awesome-agent-swarm

A curated list of AI Agent Swarm frameworks, multi-agent orchestration, swarm intelligence, and…

Andrewzh112/Awesome-LLM-based-MultiAgents

Contributors

Flashinglf

1 commits

Languages

Python

100.0%

Flashinglf/awesome-multi-agent-communication-collaboration-evolution

README

Awesome Multi-Agent Communication, Collaboration, and Evolution

Featured Technical Readings

Contents

Scope

Main Documents

Category Overview

Source Coverage

Core Papers

Core GitHub Projects and Systems

Core Technical Readings

Core Non-GitHub References

Complete Paper Catalog

Communication & Routing

Task-Oriented Collaboration & Dynamic Teaming

Debate, Consensus & Verification

Role-Playing, Agent Society & Simulation

Software Engineering Multi-Agent Collaboration

Robustness & Reliability

Evolution & Optimization

Evaluation & Benchmarks

Surveys & Taxonomies

GitHub Project Catalog

Frameworks & Runtimes

Protocols & Interoperability

Benchmarks & Evaluation

Observability & Operations

Research Implementations & Baselines

Reference Systems & Case Studies

Technical Docs, Blogs, and Field Reports

Anthropic Official Engineering Articles

OpenAI Official Engineering Articles and Docs

Multi-Agent Framework Documentation

Protocol Documentation

Evaluation, Runtime, and Engineering Blogs

High-Quality Personal and Community Blogs

Non-GitHub Project and Benchmark References

Protocols & Interoperability

Benchmarks & Evaluation

Observability & Operations

Research Implementations & Baselines

Maintenance

Contributors

Similar repos

Languages

Flashinglf/awesome-multi-agent-communication-collaboration-evolution

README

Awesome Multi-Agent Communication, Collaboration, and Evolution

Featured Technical Readings

Contents

Scope

Main Documents

Category Overview

Source Coverage

Core Papers

Core GitHub Projects and Systems

Core Technical Readings

Core Non-GitHub References

Complete Paper Catalog

Communication & Routing

Task-Oriented Collaboration & Dynamic Teaming

Debate, Consensus & Verification

Role-Playing, Agent Society & Simulation

Software Engineering Multi-Agent Collaboration

Robustness & Reliability

Evolution & Optimization

Evaluation & Benchmarks

Surveys & Taxonomies

GitHub Project Catalog

Frameworks & Runtimes

Protocols & Interoperability

Benchmarks & Evaluation

Observability & Operations

Research Implementations & Baselines

Reference Systems & Case Studies

Technical Docs, Blogs, and Field Reports

Anthropic Official Engineering Articles

OpenAI Official Engineering Articles and Docs

Multi-Agent Framework Documentation