Where should trust checks happen for AI coding agents?

I’ve been using and studying AI coding agents more, and the part I keep getting stuck on is not whether they can write code. They obviously can. The harder question is where trust is supposed to enter the workflow. If an agent touches files outside the task, skips tests, makes assumptions, or opens a PR that takes longer to review than writing it manually, the damage is already kind of done. I’m trying to understand where devs would actually want a check to happen. Before coding, when the task and allowed scope are being set? During coding, when the agent starts drifting? Before PR, so messy agent work does not enter review yet? Or during PR review, where you just need a short packet showing the original task, files touched, commands run, missing evidence, and what to inspect first? For people using Claude Code, Codex, Cursor, or custom agents in real repos: where did trust break in your last real workflow, and where would a check like this need to sit for you to actually use it instead of ignoring it as process overhead?