Handoffs are part of the agent

A handoff is part of the agent, not paperwork around the agent.

That is the part that matters when work spans more than one run. The next session should not have to rediscover state by poking around.

The useful handoff is small and concrete:

• last verified file
• next file
• current constraints
• one unfinished thing that still matters

I like handoffs that read like operator notes instead of summaries. “Checked src/routes/billing.ts and the diff is clean. Next file is src/routes/invoices.ts. Remaining issue: confirm the redirect path after login.” That is enough to keep the next run moving without forcing it to re-derive the last hour.

The other thing I want in a handoff is a clear stop point. If the last run hit a blocked assumption, say that. If a file still needs verification, say that. If a decision was made and should not be revisited, say that too. A handoff should preserve momentum, not just memory.

A good handoff also leaves one concrete verification step for the next run. For example, “run the migration test after the patch lands” is more useful than “check it later.” The next agent should be able to pick up the thread without rebuilding the whole mental model from scratch.

I usually keep the handoff in plain prose with a tiny checklist, because that is easy to scan at the beginning of the next run. If the last pass found one risky assumption and two safe files, say so directly. The point is to make the next agent useful faster, not to write a ceremony that sounds complete.