@SergioPaniego: frontier agents are this good partly because the model was trained inside the very harness it ships with great to see t…
Summary
Sergio Paniego highlights that frontier agents' performance is due to models being trained inside their deployment harness. The new work 'Polar: Agentic RL on Any Harness at Scale' by NVIDIA AI enables turning harnesses like Codex, Claude Code, Qwen Code, or Pi into RL training environments without modifying their internals.
View Cached Full Text
Cached at: 06/05/26, 09:19 PM
frontier agents are this good partly because the model was trained inside the very harness it ships with
great to see this recipe moving to the open with works like the new “Polar: Agentic RL on Any Harness at Scale” by @NVIDIAAI
it turns harnesses (codex, claude code, qwen code or pi) into RL training environments without touching their internals
Similar Articles
@billxbf: Excited to release Polar, our Agent RL rollout infra for real-world harnesses. Be it Codex, Claude Code, OpenClaw, Herm…
Polar is an agent RL rollout infrastructure that allows using real-world harnesses as training environments without code changes, supporting models like Codex, Claude Code, OpenClaw, and Hermes.
Nvidia just showed that the harness, not the AI model, is now the real hero
Nvidia's research demonstrates that a well-designed harness around an AI model, rather than the model itself, significantly improves performance on long-horizon tasks, with Claude Opus 5 achieving a perfect score on the ARC-AGI-3 benchmark.
Observation: the best agent harness for each model will be from the model developer themselves
A discussion on how AI models perform best with harnesses developed by their own creators, as third-party harnesses may cause underperformance despite strong benchmarks, citing examples like Claude Code for Claude and Codex for GPT.
OpenForgeRL: Train Harness-native Agents in Any Environment
OpenForgeRL is an open-source framework for training harness-based AI agents end-to-end in diverse environments, using a lightweight proxy and Kubernetes orchestrator to enable RL on any harness at scale. It achieves strong results on agentic benchmarks and shows that RL improves agent reliability.
@sydneyrunkle: let's assume agent = model + harness unfortunately, good models are getting really expensive! so you need a great harne…
A guide on optimizing AI agent performance by improving the harness component to compensate for expensive model costs, focusing on hill climbing techniques.