CI for AI-assisted development

How CI becomes the enforcement layer once a coding agent can produce a change faster than any reviewer can carefully read it. Covers the minimum gate (lint, type check, tests, and build, run in cost order), the contract that keeps local and CI commands aligned so an agent's fix-and-repush loop actually converges, automated screens for secrets and risky dependencies, how to stop an agent or a rushed human from weakening the gates they're checked against, why a flaky test is a bigger liability once an agent might be iterating against it, what CI feedback an agent can actually act on, and where CI's authority ends: the architectural and intent judgment that stays with human review. For builders directing coding agents who need CI to be the deterministic check a probabilistic author can't talk its way past.