The Hard Part of AI Agents Is Knowing When to Stop

An agent that can call ten tools is not automatically ready for production. The real test is what happens when permissions are missing, a dependency fails, the network times out, or the result is uncertain.nnReading, patch generation, and tests can be automated. External messages, production data changes, and payments need approval gates. Every action should retain the input version, tool arguments, result, and rollback path. Deliberately create failures and verify that the agent stops without overreach.

Sources

See the linked primary sources in the Chinese edition.