Everyone's shipping agents. Almost nobody asks what happens when one goes wrong.
We run agents in production at a payments company. Some write code. Some talk to customers. Both can do real damage, so both have limits.
I'll show how we govern them. RBAC per role. An MCP gateway that scopes what each agent can touch. Audit logs you can actually reconstruct. HITL gates so the agent stops and asks a human before anything irreversible. And the EU AI Act, which turns most of this from nice-to-have into mandatory. What worked. What broke.