A cheaper model with a better harness just beat the stronger one

Google Cloud AI Research's Harness² keeps the model frozen and rewrites everything around it: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins.

The loop proposes two edited harnesses, runs them against the current one on the same task with the same model, and keeps only the changes the traces support. Lessons get stored and retrieved per task, so one fix doesn't leak into unrelated work.

One round took GLM-5 under OpenCode from 64.2 to 75.7 on WorkBuddy-Bench, above base Gemini at 72.4.

The consequence for builders: the model is the part everyone can swap next month. The harness, and the evidence behind each piece of it, is the part that stays yours. The catch is that "better" here is behavioral evidence with no answer key, and the repo itself warns about unrestricted network access and a disabled sandbox in the Codex runs. Self-editing harnesses need controls in code.

Full breakdown, with the benchmarks and where it loses to the baseline: Google's Harness² suggests the next AI upgrade may not be a new model