Same model. Eight harnesses. Twenty-five tasks.
With the same model, success swung widely across harnesses. The harness is where teams win or lose.
Success swung from 68% to 88%.
That is not a model upgrade. That is a cage upgrade.
Marmelab just published a field audit of AI harness engineering for coding agents. They reviewed 246 repos and 57 publications. Then they filtered for harnesses that actually test themselves. Eleven made that cut.
The swing they open with is a cited same-model run: same provider, same tools, eight different scaffolds. Marmelab mapped the field. Someone else measured the cage effect. Worth keeping that attribution clean.
The consensus formula keeps showing up: Agent equals Model plus Harness. The harness is the tools, context, memory, and control loops around the model.
What the inventory shows is almost funny. Everyone writes instructions. Almost nobody enforces them or measures whether they help. AGENTS.md is winning as the shared briefing file. The craft move is turning prose into something that fires: hooks, scripts, evals.
If your agents feel random, start there. Same model. Better cage.
What is one rule in your harness you still cannot prove helps?