Agent harness lab
A deliberately badly-behaved stand-in model — it fences its JSON, drops fields, loops and fails silently, exactly like the real thing. Learn mode builds the extraction, validation, bounded retry, tool dispatch and eval layer around it; Break-fix mode fixes harnesses that swallow errors, never terminate, or ship confidently wrong answers.
Why this matters: Agent = Model + Harness, and the harness is roughly 90% of the system — which means most agent failures are configuration failures, not model limitations. These labs need no API key and no real model: the stub misbehaves deterministically, so you can build and grade the layer that actually matters, for free and reproducibly.
Harness — getting started
New to building agents? Each walkthrough gives you a deliberately badly-behaved stand-in model — it fences its JSON, drops fields, loops, and fails silently, exactly like the real thing — and you build the layer that makes it usable: extraction, validation, bounded retry, tool dispatch, and an eval harness. No API key, no bill, fully reproducible.
Loading scenarios…