Microsoft’s OpenForgeRL trains agents inside the same harnesses they use in production
dair_ai · x · 2026-07-25
Microsoft researchers and collaborators present OpenForgeRL, a framework for training agents inside the same harnesses they run in at deployment time.
- Modern agents live inside complex harnesses such as Claude Code, Codex, and OpenClaw, but most RL training happens in simplified environments that do not match production.
- OpenForgeRL uses a lightweight proxy to serve harness model calls while recording them as training data for a standard RL stack, plus a Kubernetes orchestrator that runs each rollout in its own container.
- This lets teams train directly in the real harness at scale, reducing train-deploy mismatch.
- Reported results include 72.3 on WebVoyager, 63.0 on Online-Mind2Web, and 37.7 on OSWorld-Verified, outperforming open baselines of similar size and in some cases matching much larger models.
- The paper also finds that harness choice itself is a major variable: some harnesses are much harder to learn, RL improves self-verification and multi-step completion, but error recovery remains weak.
More from coding & agent
- OpenRouter launches beta classifiers to tag and visualize model routing — AccBalanced · 2026-07-25
- Codex workflow tip: draw the flow first, then ask it to implement it — alex_frantic · 2026-07-25
- Paper argues production agents fail from context overload, not reasoning — omarsar0 · 2026-07-25
- Gemini CLI fixes Windows diff views by normalizing CRLF to LF in a2a-server — luisfelipe-alt · 2026-07-25
- VaynerX says a $20,000 internal app was prototyped in hours with Claude — tomcrawshaw01 · 2026-07-25
- Researchers open-source a 1.6B-parameter Minecraft world model and code — heghbalz · 2026-07-25