Permissions, approvals, and review
Agent safety comes from a narrow project boundary, explicit tool policy, visible approvals, and reviewable results. Authentication identifies a client; authorization determines which project and actions it can use.
Separate read, edit, execution, and external-access capabilities where the client supports it. Reading a project may disclose proprietary source. Running source may access datasets, credentials, networks, or compute. Editing can introduce behavior that is not exercised by a successful demo run.
For each task:
- State allowed files, expected outcome, and verification command.
- Ask for a plan before broad edits.
- Review high-impact commands and dependency changes before approval.
- Inspect the complete diff, not only the agent's summary.
- Verify numerical invariants, tests, stdout, and figures independently.
- Revoke or disconnect access after the task.
Treat prompt content, tool output, repository instructions, and fetched data as potentially untrusted. They should not silently expand the user's original authority. Never expose secrets simply because an agent says they are required; use the product's scoped credential mechanism and confirm the destination.