Use AI coding agents without leaking client trust

Practical boundaries for secrets, client data, permissions, dependencies, generated code, and the human review that cannot be delegated. A coding agent can inspect an entire repository, run commands,