How it works
Handoffs
Passing work between agents, not just facts.
Ordinary memory answers what is true. A handoff answers where this work stands and why, which is the thing that actually breaks when you move between threads or agents.
Five layers, all required
| Layer | |
|---|---|
| Facts | Ground truth: paths, versions, current state |
| Narrative | What happened, in sequence |
| Decisions | Choices made, and the reasoning behind them |
| Priorities | What to do next, most important first |
| Warnings | Dead ends, anti-patterns, things not to touch |
All five are required, and that is deliberate. A summary that omits why leaves the next thread to rediscover it, usually by repeating the failure that produced the warning.
Using them
Any connected tool can commit one before ending a thread and load one when starting. In practice you just ask:
Commit a handoff before we stop.
and later:
Pick up the handoff and carry on.
From code, see handoffs in the SDK.
Handoffs versus threads
They answer adjacent questions and it is worth keeping them apart:
- A thread is automatic. It records what happened, and you can reconstruct it afterwards.
- A handoff is deliberate. You are choosing what the next agent needs, including judgements that are nowhere in the transcript.
Use a thread to remember. Use a handoff to hand over.