Agent repo access scope
Your org wants to roll out an autonomous coding agent across all repos. To be 'helpful', the proposed setup gives the agent read access to every private repo (so it has full context), the ability to open PRs, and a tool that lets it run arbitrary shell commands in CI to test its changes. Several repos hold customer-data pipelines and signing keys. As the engineer reviewing this rollout, what risks do you raise and what scoping would you require?
review_agent_grants(grant_requests: list[str]) → list[str][["billing|internal|read","keys-vault|secrets|read","payments|customer_data|shell"]]out["allow","deny","deny"][["www|public|open_pr","www|public|self_merge","etl|customer_data|read"]]out["allow_with_review","deny","allow_in_vpc"]Treat the AI’s output as a draft to verify, not an answer to trust. Name the specific flaw and the input that triggers it, say how you’d catch it (tests, edge cases, reading critically), and how you’d re-prompt or decompose to get it right.
Vibe & agentic: describe the solution in plain language (or narrate it) and the coach grades your approach.