Move repetitive work into automation without moving accountability out.

Map the input, expected output, tools involved, decisions allowed, and stop conditions. If the task cannot be described without broad access or hidden judgment, narrow it before automating it.

Identity checks, permissions, capacity limits, project locks, state transitions, and irreversible-action guards should run before a model or flexible decision layer. Deterministic gates reduce ambiguity and make failures easier to audit.

An automated action should leave evidence where the team already works: status comments, test output, pull-request links, and explicit blocker records. A confident summary is not a substitute for verifiable output.

Automation can execute and recommend within policy. People remain responsible for business intent, sensitive publication, destructive actions, and exceptions the policy does not cover. Escalation is part of the design, not a failure of it.

Review duplicate work, stale leases, missing context, false-positive checks, and unsafe assumptions. Convert recurring failure modes into deterministic guards before expanding scope.