The handoff note is the only part of your agent system that survives a restart

I've run autonomous agents long enough to learn one uncomfortable truth: your logs are not your state.

An agent's log tells you what it did. It doesn't tell you what it was about to do. Restart the process — or worse, let it crash — and the agent starts from scratch, re-derives the same plan, and re-burns the same tokens. That's the retry-loop death spiral, and it's how a "cheap" agent quietly becomes a recurring bill.

The fix is a handoff note: one short file the agent writes before it stops, and reads before it starts.

The 5-part handoff that works

1. What I just finished — so the next run doesn't redo it.
2. What I was about to do — the exact next action, with the why.
3. Blockers — what's standing in the way, and whether it needs a human.
4. Open decisions — what's unresolved and who decides it.
5. The "competent stranger" test — if a stranger reads this note, can they continue the work without re-asking a single question?

The rule that makes it stick

Every run ends with a handoff, and every run starts by reading it. No exceptions. An agent that "doesn't have time" to write a handoff is exactly the agent that will restart from zero.

This is a small habit and a big difference: it's the difference between an agent that survives a crash and one that relaunches into a fresh (and expensive) loop.

I wrote this pattern down — along with the exact template and the restart-recovery flow — in the Agent Handoff Kit.