What's this about?
A framework for training agents end-to-end inside the complex harnesses they actually run in — Claude Code, Codex, OpenClaw — rather than in simplified lab environments that don't match deployment (6 upvotes).
What exactly is a "harness"?
The harness is everything wrapped around the model: the tool set, the prompts, the loop that feeds results back. Models are usually trained bare and then dropped into a harness; the mismatch between training conditions and the deployed harness is a known source of clumsiness.
Why does it matter?
It's the training-side mirror of this weekend's system-prompt story: as harnesses simplify to expose the model, the research frontier moves to making models and harnesses co-adapt — trained together, not bolted together.