AgentFlow

What it is

AgentFlow is a modular agentic framework designed to optimize tool-augmented reasoning through four specialized modules: Planner, Executor, Verifier, and Generator. It emphasizes online optimization of the planner within the system using Flow-GRPO and supports integration with multiple tools.

How it works

The framework coordinates four modules via evolving memory and an integrated toolkit to solve complex reasoning tasks across multiple turns. It includes a Flow-GRPO algorithm for in-the-flow optimization and supports multi-tool integration such as base_generator, python_coder, google_search, wikipedia_search, web_search, among others.

Getting started

Installation

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

Setup Environment Variables

Copy the .env.template to .env and edit with API keys:

• OPENAI_API_KEY
• GOOGLE_API_KEY
• DASHSCOPE_API_KEY
• TOGETHER_API_KEY
• More details in assets/doc/api_key.md

cp agentflow/.env.template agentflow/.env

Check before you run

Test Tools

cd agentflow/agentflow
bash ./tools/test_all_tools.sh

Test LLM Engines

python agentflow/scripts/test_llm_engine.py

Quick Start on Inference

python quick_start.py

Quick Start on Flow-GRPO Training

# Test environment optional steps
# Dataset preparation
python data/get_train_data.py
python data/aime24_data.py
# Flow-GRPO Training
# Create tmux session and start agentflow service (Window 0)
tmux new-session -s agentflow
bash train/serve_with_logs.sh
# Create new window (Window 1) and start training
bash train/train_with_logs.sh

Recent releases

• None

Traction

• Stars: 1989
• Forks: 233
• Open issues: 9

Behind the repo

• Not provided in the README excerpt.

Caveats

• License: MIT
• Created: 2025-09-27
• Last push: 2026-02-08
• Language: Python
• Key topics include agentic-ai, agentic-systems, llms, llms-reasoning, multi-agent-systems, reinforcement-learning, tool-augmented