Hugging Face turns Claude Code, Codex and 10 coding harnesses into open RL environments
huggingface · x · 2026-10-05
Hugging Face CEO Clément Delangue announced that Claude Code, Codex, Hermes, Pi, opencode and other coding harnesses can now be used directly as RL training environments — no harness or training code changes, fully open source. Key insight: the same model scores 62% under Mini-SWE-Agent but only 33% under Claude Code, meaning most RL training happens in scaffolds nobody ships. The fix is a capture proxy speaking the 4 major agent API formats (OpenAI Chat Completions/Responses, Anthropic Messages, Gemini), forwarding to vLLM, recording exact token IDs and logprobs, and handing trainable sequences to TRL. Ten harnesses already run through it.
More from coding & agent
- Zero Game-Dev Experience: He Built a Game Boy Game for His 9-Month-Old With Codex — Hacubu · 2026-10-05
- Chris Lattner to talk Mojo type system, dependent types and comptime at SPLASH conference — clattner_llvm · 2026-10-05
- Claude Code can now read full docs via NotebookLM MCP without spending a token — Aiden_Tech_Ai · 2026-10-05
- Jacky Zhao leaves Replit to join Anthropic — Kyrannio · 2026-10-05
- Chinese agent swarms discovered in the wild, new report documents the fleet — Puzzleheaded-King584 · 2026-10-05
- OpenAI Codex Pledges Daily Improvements for 28 Days or Users Get Free Resets — udmrzn · 2026-10-05