Compose.   Adapt.   Evolve.

Compose the Harness, define the Agent.
From zero-code to full customization — one core, X entry points.

Homepage • Overview • Architecture • Quick Start • Benchmarks • Roadmap • 中文文档

🔭 Overview

The harness — not just the model — determines agent performance. The same base model produces dramatically different results depending on how context is managed, how tools are orchestrated, how errors are recovered, and how evaluation signals feed back.

HarnessX is a harness foundry: forge any number of agent harnesses from reusable processors and bundles, pair each with any model, and evolve them through training — all without rewriting the agent.

Most frameworks solved model swapping. Behavior swapping remains expensive — switching from a coding agent to a research agent, adding memory or guardrails, means rewriting the agent.

HarnessX solves this with one clean separation:

agent = model.agentic(harness)

• ModelConfig — provider routing, fallback, per-role model assignment
• HarnessConfig — the full behavior pipeline (tools, memory, processors, trace, sandbox)

The X in HarnessX stands for eXtensible Behavior Composition — compose, adapt, and evolve harnesses without rewriting the agent:

🧩 Compose — 9-dimension behavior pipeline; any behavior = Processor, combine with | operator.

⚙️ Adapt — Harness observes performance and auto-searches optimal harness configurations.

🚀 Evolve — every run produces reward-annotated trajectories that feed SFT / RL training.

🏗️ Architecture

→ See docs/architecture.md for the full 9-dimension behavior pipeline, processor hook points, and composition API.

🚀 Quick Start

curl -sSf https://raw.githubusercontent.com/Darwin-Agent/HarnessX/main/scripts/install.sh | bash

curl -sSf https://raw.githubusercontent.com/Darwin-Agent/HarnessX/main/scripts/install.sh | bash -s -- --all

uv python install 3.12
uv venv --python 3.12 .venv
source .venv/bin/activate
uv pip install -e .
# Build frontend (required for hx lab)
cd frontend && npm install && npm run build && cd ..

export ANTHROPIC_API_KEY=sk-...

hx "Research 2026 AI agent trends and write a structured report"
hx -p "Write a Python fizzbuzz"     # non-interactive, print and exit
hx -c path/to/config.yaml           # load a YAML config
hx --resume <run_id>                # resume a previous session
hx lab                              # open the Lab UI at localhost:8000

hx-gateway start   # start the gateway (configured in ~/.harnessx/gateway.yaml)

import asyncio
from harnessx import BaseTask, HarnessConfig
from harnessx.core.model_config import ModelConfig
from harnessx.providers.anthropic_provider import AnthropicProvider

async def main():
    model = ModelConfig(main=AnthropicProvider("claude-sonnet-4-6"))
    harness = model.agentic(HarnessConfig())
    result = await harness.run(BaseTask(description="What is 2 + 2?"))
    print(result.final_output)

asyncio.run(main())

📊 Benchmarks

HarnessX provides two evolution loops that systematically improve agent performance on any benchmark:

• Harness Evolution — a meta-harness analyzes trajectories and automatically searches for better processor combinations, prompt strategies, and tool configurations, without changing the model.
• Model Evolution — reward-annotated trajectories from harness runs feed RL fine-tuning (via VERL), improving the model itself.

The two loops compose: evolve the harness first, then evolve the model on top. Below are results on the GAIA benchmark. See benchmarks/README.md for additional benchmarks and adapter details.

Harness Evolution (Qwen 3.5 9B)

Starting from a default harness (R0, 33%), the meta-harness discovers better configurations round by round — reaching 47% by R3, a +14pp gain with zero model changes. → Reproduce: recipe/gaia_evolver/

Harness Evolution (GPT-5)

The same approach scales to frontier models. Overall GAIA accuracy rises from 62% to 84% after evolution, with gains across all five domains. → Reproduce: recipe/gaia_evolver/

Model-Harness Co-Evolution (Qwen 3.5 9B)

When the two loops run together, the gains compound: harness evolution lifts the baseline from 33.97% to 41.67%; model evolution pushes it further to 55.77% — a +64% relative improvement, all on a 9B model. → Reproduce: recipe/verl_harnessX/

📁 Project Structure

HarnessX/
├── harnessx/                  # 🧠 Core framework
│   ├── core/                  #    Harness, Builder, RunLoop, State, Events, Trajectory
│   ├── processors/            #    7 categories × multiple processors
│   │   ├── context/           #    📝 System prompt, history, user wrapper
│   │   ├── control/           #    🛡️ 13 safety & reliability processors
│   │   ├── evaluation/        #    📊 LLM judge, PRM, self-verify
│   │   ├── memory/            #    🧠 Extraction, retrieval, 5 strategies
│   │   ├── multi_model/       #    🔗 Model routing
│   │   ├── observability/     #    🔭 OTel, checkpoints, metrics
│   │   └── tools/             #    🔧 Skill loader, schema adapter, filters
│   ├── providers/             # 🔌 6 model backends + agentic mixin
│   ├── plugins/               # 🧩 Plugin base, discovery, builtins, dimensions
│   │   └── dimensions/
│   │       └── light_memory/  # 🧠 Light-Memory (self-developed)
│   ├── tools/                 # ⚒️ Tool registry, builtins
│   ├── sandbox/               # 📦 Local, Docker, E2B
│   ├── tracing/               # 📡 Journal, OTel, null tracer
│   ├── rl/                    # 🧬 RLConfigSpec, TaskBuilder
│   ├── bundles/               # 📦 Pre-composed capability bundles
│   ├── api/                   # 🌐 FastAPI + SSE for Lab UI
│   └── cli.py                 # ⌨️ CLI entry point (hx)
├── benchmarks/                # 📊 4 integrated + 3 ongoing benchmarks
├── recipe/                    # 🧪 slime (RL training recipe)
├── examples/                  # 📖 coding / research / assistant / custom_processor
├── extensions/                # 🔌 Skills (docx, pdf, pptx, xlsx)
├── frontend/                  # 🖥️ Lab UI (React + TypeScript + Tailwind)
└── tests/                     # ✅ Unit, integration, E2E

🗺️ Roadmap

For detailed design notes and motivation behind planned items, see ROADMAP.

Phase	Focus	Status
1	Core: 9-dimension behavior pipeline, 13 processors, multi-provider, SFT/RL bridge, 4 benchmarks, Lab UI
2	Meta-opt: Bayesian Optimization, Meta-Harness, auto config search
3	Self-evolution: closed-loop training, HarnessHUB community marketplace
4	Memory: multimodal backends, third-party integrations (VERL, SuperMemory, OpenVKing)

In-Repo Implementations

• Light-Memory — file-based memory with time-decay, daily compression, git versioning (harnessx/plugins/dimensions/light_memory/)
• Slime RL recipe — SGLang rollout adapter + token annotation + GRPO training pipeline (recipe/slime/)
• MetaHarness — agent observes its own trajectories and proposes harness config changes; observer harness + meta-agent + sandboxed promotion loop
• LoCoMo benchmark — long-context memory evaluation: session recall, cross-turn consistency, compaction fidelity
• Bayesian Optimization — surrogate model search over the ~10^6-configuration harness space
• HarnessHUB — community platform to publish, version, and pull HarnessConfig bundles (hx pull [email protected]; Lab UI panel; private registries)
• Multimodal Memory — CLIP-based image/video memory backend via the plugin system
• Harness Memory Evolution — closed loop: trajectories → RL fine-tuning → better model → better harness; population-level mutation + data flywheel

Third-Party Integrations (opt-in, live in recipe/)

• VERL — connect HarnessX rollouts to distributed PPO / GRPO training loops
• MemPalace — structured episodic memory backend
• SuperMemory — cloud-backed semantic memory via the plugin system
• OpenVKing — vector-knowledge-graph memory for entity-rich domains
• Memory quality metrics — retrieval precision / recall surfaced through HarnessJournal
• Data synthesis pipeline — controlled SFT / preference-dataset generation with diversity constraints

🤝 Contributing

HarnessX is fully open-source under the MIT License. Contributions are welcome for:

• 🧩 New processors — behavior modules for unexplored dimensions
• 🧠 New memory backends — via the plugin system
• 📊 New benchmark adapters — benchmarks/ pattern
• 🧪 RL training recipes — recipe/
• 🖥️ Lab UI improvements

Please read CONTRIBUTING.md first.

@software{harnessx2026,
  title   = {HarnessX: A Composable, Self-Evolving Agent Harness Foundry},
  author  = {Darwin Agent Team},
  year    = {2026},
  url     = {https://github.com/Darwin-Agent/HarnessX},
  license = {MIT},
}