AgentFlow

Frameworks & SDKs 💻 Python ⚖️ MIT 🔴 No recent commits
2.0k stars

AgentFlow is a trainable, tool-integrated agentic framework that optimizes a modular system of Planner, Executor, Verifier, and Generator agents using Flow-GRPO. It solves scalability and generalization limits in tool-augmented reasoning for long-horizon tasks. Aimed at developers and researchers building advanced AI agents.

7B Agent Outsmarts a 200B LLM: AgentFlow by Stanford

🎬 7B Agent Outsmarts a 200B LLM: AgentFlow by Stanford · Discover AI

✨ Key features

  • Modular agentic system with four specialized modules
  • Multi-tool integration including search, code, and generators
  • Flow-GRPO algorithm for in-the-flow agent optimization
  • Proven results on 10 benchmarks, outperforming GPT-4o
  • Supports custom LLM engines for each agent module

🎯 Use cases

  • Complex question answering with web search
  • Mathematical reasoning with tool use
  • Agentic reasoning tasks requiring multi-step planning
  • Benchmarking agentic systems on search, math, and science tasks

📦 Installation

🧰 Requirements: Python 3.11, API keys for OpenAI, Google, and optionally DashScope or Together, or serve a local vLLM model.

bash setup.sh
source .venv/bin/activate
(Optional) Install `parallel` for running benchmark experiments in parallel:
sudo apt-get update
sudo apt-get install parallel

Copy the .env.template file from agentflow/.env.template and rename it to .env, then place it in the agentflow/ folder. Update the following variables with your own API keys:

  • OPENAI_API_KEY (for judging reasponse)
  • GOOGLE_API_KEY (for Google Search tool)
  • DASHSCOPE_API_KEY ([optional] for calling Qwen-2.5-7B-Instruct as engine for agents and tools)
  • TOGETHER_API_KEY ([optional] alternative for calling Qwen-2.5-7B-Instruct as engine for agents and tools - recommended for international users)
  • More ways: serve Qwen2.5-7B-instruct model with vLLM (details refer to serve_vllm_local.md).
cp agentflow/.env.template agentflow/.env
# Then edit agentflow/.env with your API keys

🚀 Usage

python quick_start.py

Example output:

==> Initializing agentflow...
==> Setting up tools...
==> 🎯 Reasoning Steps from AgentFlow (Deep Thinking...)
==> 🔍 Step 0: Query Analysis
==> 🎯 Step 1: Action Prediction (Google_Search_Tool)
==> 🛠️ Step 1: Command Execution (Google_Search_Tool)
...
**Answer:** The capital of France is Paris.
==> ✅ Query Solved!

❓ FAQ

What are the prerequisites for installing AgentFlow?

Python 3.11 is recommended. You also need to set up API keys for OpenAI, Google, and optionally DashScope or Together, or serve a local model with vLLM.

How do I test if my environment is correctly configured?

Run bash ./tools/test_all_tools.sh to test tools and python agentflow/scripts/test_llm_engine.py to test LLM engines.

Can I use my own model with AgentFlow?

Yes, you can modify the llm_engine_name parameter in run.sh for the planner, and for other agents modify self.llm_engine_fixed in planner.py and the Executor instantiation in solver.py.

How do I train AgentFlow with Flow-GRPO?

Prepare datasets with python data/get_train_data.py and python data/aime24_data.py, then start training using tmux with bash train/serve_with_logs.sh and bash train/train_with_logs.sh.

📊 Repository

Stars★ 2,028
Forks🍴 235
Open issues🐛 9
Last commit🕒 Feb 8, 2026
Created📅 Sep 2025
Language💻 Python
License⚖️ MIT

🤖 Overview, features, install steps and FAQ were generated from the project's README on Sep 4, 2026. Always check the original source before running commands.