AgentFlow: In-the-Flow Agentic System Optimization

AgentFlow is a trainable, modular agentic framework designed to overcome the scalability and generalization limits of current tool-augmented reasoning approaches. It targets researchers and developers building sophisticated AI agents, offering improved planning, tool use, and long-horizon reasoning capabilities through direct optimization of its planner module.

How It Works

The framework employs a modular system comprising four specialized agents: Planner, Executor, Verifier, and Generator, which coordinate via evolving memory and integrated tools. AgentFlow's core innovation is the Flow-GRPO algorithm, which optimizes the planner agent "in-the-flow" online. This approach is advantageous for tackling long-horizon reasoning tasks with sparse rewards, enhancing tool-calling reliability and generalization beyond single-LLM interleaving methods.

Quick Start & Requirements

• Installation: Execute bash setup.sh. Activate the virtual environment with source .venv/bin/activate. Install parallel for benchmark experiments: sudo apt-get install parallel.
• Environment Variables: Configure API keys (OpenAI, Google, DashScope/Together) in agentflow/.env by copying agentflow/.env.template. Detailed instructions are available in the API Key Setup Guide.
• Inference: Run python quick_start.py after setting up API keys.
• Training: Requires dataset preparation (python data/get_train_data.py, python data/aime24_data.py) and execution via tmux using train/serve_with_logs.sh and train/train_with_logs.sh. Hyperparameters are configured in train/config.yaml.
• Documentation: Links to a YouTube tutorial, test_env.md, benchmark.md, logs.md, and llm_engine.md (relative path assets/doc/llm_engine.md) are available within the repository structure.

Highlighted Details

• AgentFlow (7B Backbone) achieves state-of-the-art results on 10 benchmarks, outperforming baselines by up to +14.9% (search) and surpassing GPT-4o (~200B) on several tasks.
• The core innovation is the Flow-GRPO algorithm, enabling "in-the-flow" online optimization of the planner agent for long-horizon reasoning and sparse rewards.
• Features a modular agentic system (Planner, Executor, Verifier, Generator) with evolving memory and seamless integration of diverse tools (Python, Google Search, Wikipedia, etc.).
• Supports customization by allowing different LLM engines for each agent module.

Maintenance & Community

Core contributors include Zhuofeng Li, Haoxiang Zhang, and Pan Lu, with advisors James Zou, Yejin Choi, and Yu Zhang. A Slack community is available for contributions and discussions. Contact emails are provided for issues and collaborations.

Licensing & Compatibility

The project's license is not explicitly stated in the provided README, which is a critical omission for adoption decisions. The framework supports various LLM engines for its agent modules.

Limitations & Caveats

The setup requires obtaining and configuring multiple API keys, potentially incurring costs. The absence of a clearly stated license poses an adoption blocker. Training involves a multi-step process using tmux and specific dataset preparation.