AgentJet

AgentJet (AJet) is a cutting-edge, user-friendly agent RL training framework designed to optimize agents and agentic workflows (supporting any agent built with OpenAI SDK, AgentScope, Langchain, or raw HTTP requests), fine-tuning LLM weights to enhance model performance.

AgentJet (AJet) has fully-distributed swarm training capability, which means that you can deploy ajet-swarm start in GPU server(s) and then start training agents in your laptop(s)! Simply provide your agent workflow, training dataset, and reward function, and AgentJet will be ready to go!

✈️ News

• 2026.6.4 Arxiv paper release: AgentJet - A Flexible Swarm Training Framework for Agentic Reinforcement Learning 2606.04484.
• 2026.5.26 From Laptop to GPU Cluster: sync_train_code enables remote-controlled Swarm Training and closes the loop for automated RL research. EN Blog / ZH Blog.
• 2026.4.30 DeepFinance: training a financial deep-research Agent with multi-dimensional reward GRPO. EN Blog / ZH Blog.
• 2026.4.9 AgentJet-powered Auto Research: fully automated overnight RL experiments. EN Blog / ZH Blog.
• 2026.3.30 LoRA training example is now online! See tutorial/example_math_lora for an example.
• 2026.3.26 Upgrade verl backend to 0.7.1 to support more models and increase training speed! All benchmark verified.
• 2026.3.19 Support for latest Qwen3.5 models is in progress.
• 2026.3.12 Tuning Original OpenClaw Agent without Editing Any Agent Code. EN Blog / ZH Blog.
• 2026.3.09 Non-shared-parameter Multiagent Training. EN Blog / ZH Blog.
• 2026.2.20 Introducing AgentJet Swarm. ZH Blog / EN Blog.

✈️ Fast Introduction

1 - Classic Mode

Let's begin with the simplest example: a math agent with a tool call. This is a simple & centralized training example.

1. please check out the installation guide to set up the training environment.
2. tune your first model using the minimum example.

   ajet --conf ./tutorial/example_math_agent/math_agent.yaml --backbone='verl'
   

2 - Swarm Mode

Let's begin with the simplest AgentJet Swarm example: also a math agent. In this case, you can use any GPU-less laptop to train the model remotely.

1. Start swarm server and begin swarm overwatch: ajet-swarm start and ajet-swarm overwatch (or ajet-swarm top). (Alternative: if you are a fan of docker, use our prebuilt docker image here without setting up dependencies)
2. From your laptop (or swarm server localhost), run this simple script to begin training:

   AJET_SWARM_URL="http://swarm-server-ip:10086" python ./tutorial/example_math_swarm/math.py
   

✈️ Features

We aim to build an easy-to-learn Agent tuner that unlocks more possibilities for agent developers:

• Easy and Friendly. AgentJet helps you tune models behind your agent workflows easily, optimizing your agents for top performance with minimal effort.
• Rich Tutorial Library. AgentJet provides a rich library of examples as tutorials.
• Swarm Training. This unique feature of AgentJet opens many possibilities: deploying distributed & self-healing rollout workers, non-shared-parameter multi-agent training, multi-runtime & multi-task cocktail training. And just like Tinker, you can use AgentJet Swarm to train models even on GPU-less laptop(s).
• Efficient and Scalable. AgentJet uses [verl] as the default backbone (--backbone=verl). However, we also support trinity as an alternative backbone, accelerating your tuning process via fully asynchronous RFT.
• Flexible and Fast. AgentJet supports multi-agent workflows and adopts a context merging technique, accelerating training by 1.5x to 10x when the workflow involves multi-turn (or multi-agent) conversations.
• Reliability and Reproducibility. Our team keeps track of framework performance across multiple tasks + major-git-version + training-backbones (under construction, still gathering data, coming soon).

For advanced researchers, AgentJet also provides high-resolution logging and debugging solutions:

• High-Resolution Logging: AgentJet allows users to save and inspect token-level rollout details, recording token IDs, token loss masks, and even token logprobs to facilitate workflow development and agent diagnostics.
• Fast Debugging: AgentJet also provides the --backbone=debug option for the best debugging experience, shortening your wait period from minutes to seconds after code changes and enabling breakpoint debugging in IDEs.

---

✈️ Quick Start

Installation

• Click here to read the installation guide.

Example Library

Explore our rich library of examples to kickstart your journey:

• 🔢 Training a math agent that can write python code.
• 📱 Creating an AppWorld agent using AgentScope and training it.
• 🐺 Developing Werewolves RPG agents and training them.
• 👩🏻‍⚕️ Learning to ask questions like a doctor.
• 🎴 Writing a countdown game using AgentScope and solving it.
• 🚶 Solving a frozen lake walking puzzle using AgentJet.

Explore our automated benchmarking system https://benchmark.agentjet.top/:

---

✈️ Core Concepts

AgentJet makes agent fine-tuning straightforward by separating the developer interface from the internal execution logic.

1. The User-Centric Interface

To optimize an agent, you provide three core inputs:

• Trainable Workflow: Define your agent logic by inheriting the Workflow class, supporting both simple agent setups and advanced multi-agent collaborations.
• Task Reader: Load training tasks from JSONL files, HuggingFace datasets, interactive environments, or auto-generate them from documents.
• Task Judger: Evaluates agent outputs and assigns rewards to guide training.

2. Internal System Architecture

The internal system orchestrates several specialized modules to handle the complexities of RL training and agent interactions.

• Launcher: Manages background service processes (Ray, vLLM) and routes the backbone.
• Task Reader: Handles data ingestion, augmentation, and filtering.
• Task Rollout: Bridges LLM engines and manages the Gym environment lifecycle.
• Task Runner: Executes the Agent workflow and calculates rewards.
• Model Tuner: Forwards inference requests from the workflow to the LLM engine.
• Context Tracker: Monitors LLM calls and automatically merges shared-history timelines to improve training efficiency by 1.5x to 10x.
• Swarm Server: A data interchange center that accepts OpenAI-like requests and engine instructions, activated only in AgentJet Swarm mode.

3. Swarm Architecture

When swarm training mode is enabled, an additional component will be activated:

• Swarm Data Interchange Server: Maintains HTTP service, listens to swarm instructions and OpenAI compatible requests. Establishes a high-speed zmq communication channel to coordinate other modules.

✈️ Navigation

• Tutorials: From Installation to Tuning your first agent to Multi-Node Training — the essential path for beginners.
• Core Components: Define your Trainable Workflow and manage Data and Reward.
• Example: Check the Example Library above for real-world cases like Math, Werewolves game and Learning to ask task.
• Deep Dive: Master advanced Configuration.

✈️ Roadmap

AgentJet is a constantly evolving project. We are planning to add the following features in the near future.

✈️ Citation

If you use AgentJet in your research, please cite:

@software{
  title  = {AgentJet: A Cutting-Edge Multi-Agent Training Platform for Large Language Models.},
  author = {The AgentJet Team},
  url    = {https://modelscope.github.io/AgentJet/},
  month  = {01},
  year   = {2026}
}

---

---