a-hat-optimizer boosts agent tool calling

This is a clever agents hack because it promises much better tool use without finetuning, but the big gains are still self-reported and need outside replication.

• –If the signal generalizes, it suggests models often internally know a tool call is needed before their token output reliably says so.
• –Smaller open-weight models look like the biggest beneficiaries, which could make cheap agent stacks much more usable for search, code, and file workflows.
• –The package is practical rather than purely theoretical, bundling contrastive extraction, calibration strategies, and runtime prediction into a lightweight Python library.
• –The obvious caveat is validation: there is no broad independent benchmark yet, so developers should treat the +26-58% improvement claims as promising, not settled.