#5252nd of 53 in Training and fine-tuning
agentflow
Agent system with an RL-trained planner, executor, verifier and generator
- Stars
- 2.1k
- License
- MIT
- Last commit
- Jan 2026
- Language
- Python
1,718 on 21 Apr 20262,091 on 10 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
AgentFlow is a research framework that splits an agent into Planner, Executor, Verifier and Generator modules sharing an evolving memory and a toolkit (python_coder, google_search, wikipedia_search, web_search). The planner is trained online with Flow-GRPO on verl, using NQ and DeepMath-103K. The repo includes inference scripts, benchmark runners and a 7B planner model on Hugging Face.
Who it is for: Researchers training tool-using agents with RL
Strengths
- Trains only the planner in-system with Flow-GRPO; hyperparameters are in train/config.yaml
- Each module can use a different engine: OpenAI, DashScope, Gemini, DeepSeek, Together or local vLLM
- Ships a 7B planner checkpoint, benchmark scripts and a Gradio demo
- Paper accepted at ICLR 2026; README reports gains on 10 benchmarks
Weaknesses
- Changing the Executor, Verifier or Generator engine requires editing source files
- Needs several API keys (OpenAI for judging, Google for search) even for basic runs
- No tagged release and no Docker setup; install is via setup.sh
- Research code with no general-purpose deployment or UI beyond the demo
What it needs
- GPU optional
- Needs OpenAI API, Google Search API, DashScope or Together API, vLLM, verl, tmux
- Models: Qwen2.5-7B-Instruct, AgentFlow planner 7B, GPT-4o, Gemini, DeepSeek
Quick answers
Is agentflow open source?
Yes. Its code is on GitHub under the MIT license.
Does agentflow need a GPU?
No. It runs on a CPU and can use a GPU when you have one.
Also in Training and fine-tuning
See all 53| Rank | Project | Score |
|---|---|---|
| 1 | unslothDesktop and web app to run and fine-tune local models | 88 out of 100 |
| 2 | peftPython library for LoRA and other parameter-efficient fine-tuning of Transformers models | 88 out of 100 |
| 3 | deepspeedPyTorch library for distributed training of very large models with ZeRO | 86 out of 100 |
| #4 | verlReinforcement learning training library for LLM post-training with PPO and GRPO | 86 out of 100 |
| #5 | trlPost-training library with SFT, DPO, GRPO and KTO trainers | 86 out of 100 |