Autoresearch plugin brings Karpathy loop to Claude Code

The point here isn’t a flashy score jump; it’s that the agent spent its failures buying certainty. A 93% failure rate is a feature when each revert narrows the search space and tells you where to stop tuning.

• –Ranking, not recall, was the bottleneck, so larger candidate pools and title matching were mostly dead ends.
• –The adaptive weighting survived ablation, which is exactly the kind of “we should simplify this” assumption autoresearch can settle fast.
• –Round 2 showed a classic co-optimization trap: prompt changes broke weights, and stale cache keys masked the effect until the caching bug was found.
• –This only works when the eval path is cheap, deterministic, and noisy inputs like temperature are pinned down.
• –The broader win is workflow, not optimization magic: let Claude explore the edges overnight, then spend human time on the architectural ceiling, not manual guesswork.