Serving Infrastructure: Queues, Rate Limits, and Rainbow Deployments

Suppose this happens to Desk's team. At 2:00 in the afternoon, a Desk run stops at step 6 to wait for an employee's approval of a refund. Its saved state records that it has requested a refund with a particular tool definition. At 3:00, the team releases a new version that renames that tool. At 5:00, the employee approves, and the run resumes on the new code. It fails, because its tool no longer exists.

Nothing in the agent's design caused this, and no model made a mistake