OpenForgeRL trains harness-native agents inside Claude Code, Codex, and OpenClaw rather than in stripped-down RL sandboxes

arXiv·high signal

A new arXiv paper from Xiao Yu, Baolin Peng, and Ruize Xu argues that modern agents are inseparable from their inference harnesses — the multi-turn reasoning, tool routing, and environment access provided by Claude Code, Codex, and OpenClaw — and that training agents outside those harnesses produces a train/serve mismatch. OpenForgeRL proposes training policies directly against the harness itself in arbitrary environments, treating the scaffolding as part of the learned system. For builders running agents inside a CLI harness daily, this is the first serious framing of 'the harness is the environment,' which implies model fine-tuning and harness design can no longer be decoupled.