HarborRL open-sources agentic RL: 200+ tasks, 40+ harnesses, one reproducible pipeline
Dear_Treat3688 · reddit · 2026-10-27
The datamllab team released HarborRL, an open-source framework aiming to democratize agentic reinforcement learning by unifying model serving, trajectory generation, tool execution, verifier rewards, and policy updates into a single reproducible pipeline.
Key features (Models × Algorithms × Tasks × Harnesses):
- Models: Qwen and GLM support
- Algorithms: GRPO and DAPO
- Tasks: 200+ tasks including Terminal-Bench, SWE-Bench, and Deep-SWE
- Harnesses: 40+ agent frameworks including Claude Code, Codex, LangGraph, and SWE-agent
The project is brand new (17 stars on GitHub) and still in early stages.
More from coding & agent
- Filter-then-locate beats whole-file prompts: testing 4 ways to update READMEs with LLMs — KhuyenTran16 · 2026-10-27
- Qiaomu Clipper launches site: open-source Chrome clipper that saves reading notes to Obsidian — vista8 · 2026-10-27
- Author curates 369 open-source repos for AI video editing with Claude Code and Codex agents — dotey · 2026-10-27
- Weekend vibe coding project tracks iPhone Duo pickup stock across 29 NYC Apple Stores every 5s — MichaelRoyzen · 2026-10-27
- Qdrant benchmarks rerankers across 5 datasets: confirmed gains on only 2, tune fusion first — qdrant_engine · 2026-10-27
- Meshy joins MCP for Blender: generate 3D game assets in Blender with any LLM — sidahuj · 2026-10-27