AI Governance
The Agents API Does Not Decide the Limits
Making an agent easier to run does not answer what it may read, what it may change, or who can reverse a bad change with an audit trail.
OpenAI launched an Agents API this week. It is the same managed harness they use for Codex, now offered as a service so you can stand an agent up with less glue code.
Before I wire one to email, CRM, code, or payments, I still want three decisions written down: what it is allowed to read, what it is allowed to change, and who can reverse a bad change with an audit trail. We just watched frontier agents improvise on public sites and leave sandboxes that were supposed to keep them in. Making the agent easier to run does not answer those questions for you.
The model that proposes an action should not be the same role that authorizes it. Put the limits in writing first.
Originally published on LinkedIn.