Agent Training Kit

Agent Training Kit

Agent Training Kit lets developers build trainable Strands Agents in just a few lines of code.

Agent Training Kit (ATK) makes [Strands Agents](https://github.com/strands-agents/sdk-python) trainable — agents learn from their tool usage and task execution results, improving through experience. ATK adapts [rLLM](https://github.com/rllm-org/rllm) for Strands Agents, handling agent execution, trajectory collection, and reinforcement learning.


*Figure 1: Agent Training Kit training workflow and data flow*

Code structure

Python file names match the component labels shown in Figure 1.

├── agentboost
│   ├── agents
│   │   ├── strands_agent.py        # strands agent wrapper
│   │   ├── strands_process_pool.py # parallel process execution of agents
│   │   └── strands_worker.py       # agent worker managed by the process pool
│   ├── environments
│   │   └── strands_env.py          # RL environment for strands agent             
│   └── tools
│       └── strands_code_tool.py    # tools for executing the code
├── apply_patches.py                # install dependencies 
├── examples
│   └── deepcoder
│       ├── prepare_deepcoder_data.py       # prepare dataset for training
│       ├── train_deepcoder_with_strands.py # training python script
│       └── train_deepcoder_with_strands.sh # training bash script
├── modified_rllm_files
│   ├── agent_execution_engine.py   # rLLM execution engine for training agents
│   └── agent_ppo_trainer.py        # rLLM agent trainer
├── modified_verl_files
│   └── vllm_async_server.py        # rLLM/veRL vLLM async server
└── setup.py                        # setup python modules

Installation

curl -LsSf https://astral.sh/uv/install.sh | sh

uv venv ~/venv/agentboost --python=3.12
source ~/venv/agentboost/bin/activate

git clone https://github.com/rllm-org/rllm
cd rllm
git checkout 1fc3c4babfe9a63d809d6bf9a9011df777f30c91
git submodule upd