2026-09-21-eve-auto-a-model-can-now-approve-tool-calls

auto(): a model can now approve tool calls. A fourth approval helper between once() and always(): routine calls run without a human, and a classifier decides what is routine.

Channel: tagged-release. Half: capability. Date: 2026-09-17.

Operator consequence: Try it only where a wrong “clear” is recoverable. The classifier sees the tool name and input, not the downstream effect, and tool input leaves your deployment for the evaluator’s provider. It answers the contract question “does the approval-gate model change how teams gate tool calls?”: yes, eve now ships a first-party model-judged gate, which moves human attention from every call to the classifier’s false negatives.

Receipt

• https://github.com/vercel/eve/blob/eve%400.63.0/docs/tools/human-in-the-loop.md