ART
A tool for training multi-step agents to perform real-world tasks using GRPO.
- Region
- Overseas
- Pricing
- Free
- Open source
- Yes
- GitHub Stars
- ★ 9.9k
- Source
- GitHub
- Added
- 2026-06-04
- Last verified
- 2026-06-04
- GitHub
- github.com/OpenPipe/ART
Overview
Agent Reinforcement Trainer (ART) is a tool that uses GRPO to train multi-step agents for real-world tasks. It enables agents to better complete complex tasks by training them in actual working environments. ART supports multiple models including Qwen3.6, GPT-OSS, and Llama, making it suitable for various scenarios where agent performance needs to be enhanced through reinforcement learning.
Key features
- ▪Uses the GRPO algorithm
- ▪Supports multiple models
- ▪Designed for real-world tasks
- ▪Provides online training
Use cases
Pros
- ＋Supports multiple models
- ＋Suitable for complex tasks
- ＋Improves agent performance
Limitations / notes
- －Requires some technical background
Who it's for
This overview was compiled by AI from public sources and may contain inaccuracies — please refer to the official site.
FAQ
Which models does ART support?
ART supports multiple models including Qwen3.6, GPT-OSS, and Llama.
Similar agents
Cursor
FeaturedAI-powered code editor with Agent mode
Dify
FeaturedOpen-source LLMOps + Agent orchestration platform
扣子 Coze
FeaturedByteDance's no-code Agent/Bot building platform
Trae
FeaturedAI IDE (with Agent) from ByteDance
LangChain
FeaturedThe most mainstream LLM/Agent orchestration framework
Devin
FeaturedAutonomous software engineer Agent by Cognition
Something wrong? Let us know on the About page and we'll fix it.