What a harness is ## What a harness is around an agent model In this channel's material, a harness is the agent scaffolding that surrounds an underlying model — the client-side software (think Claude Code, Codex, or similar agent environments) through which a model actually does its work. The model alone doesn't determine performance; the harness it's run inside matters just as much. When the Every team evaluates new models, harness choice is treated as part of the evaluation itself: harnesses performed drastically differently on evals, so a model's results can't be read apart from the harness they came from [^4872db36-e98a-5f6b-bc5c-b11e48c9411b@813]. Why harness and model are paired ## Harness and model are very paired A recurring theme is that you cannot hot-swap models into any harness and expect the same results. Katelyn Lesse notes that new models require harness and model to be very paired — tuning them together is what makes an agent setup perform well [^4872db36-e98a-5f6b-bc5c-b11e48c9411b@692]. That's why the channel's model vibe checks treat the harness as a caveat to understand before reaching for a model like Opus 4.8: its measured behavior reflects the harness it ran in, not the raw model alone. Harness engineering vs. the real wall ## Harness engineering — and the harder wall Teams often expect harness engineering to be the bottleneck in shipping agents. Katelyn Lesse argues the real wall everyone hits is production infrastructure — keeping servers running, storing transcript data, and secure sandboxing [^4872db36-e98a-5f6b-bc5c-b11e48c9411b@1181]. The harness matters, but it isn't where most teams should expect to get stuck. - A harness is the agent scaffolding around a model — the software layer through which the model acts. - Harnesses performed drastically differently on evals, so model results depend on the harness used. - New models need harness and model to be very paired, not hot-swappable. - Teams expect harness engineering to be the bottleneck, but production infrastructure is the real wall.