Agents need a handover protocol

An agent that works for minutes or hours does not merely need a better prompt. It needs a handover protocol: a durable account of what it was asked to do, what it tried, what remains uncertain and who can take over.

OpenAI’s new Agents API makes long-running work, tool use and subagent coordination part of a managed platform. The shift is practical, but it also exposes a basic institutional problem. Delegated work cannot remain legible only inside a model context window. OpenAI’s Agents API announcement

Continuity is a control

A handover is not a summary written for convenience. It is a control surface. It should preserve the task boundary, current state, tool calls, artifacts, open questions and stop conditions. Without that record, the next agent or human has to infer authority from leftovers.

This matters whenever work crosses sessions, models or operators. A system that can delegate can also lose the reason a permission was granted. The longer the task, the more important it becomes to distinguish completed work from plausible-looking intermediate output.

The mature agent system will treat handover as part of execution, not an afterthought. It will make continuity inspectable, interruptible and boring. Boring is a respectable achievement when software is acting on someone else’s behalf.