In a relay race, the baton is not the runner’s memory of the race so far. It is a small, standardized object designed to be passed at speed, and the race is won or lost in the exchange.
Sub-agent notes are batons. When the driller hands the session to the evaluator, the transcript of the drilling phase does not transfer; a note does. The departing agent keeps taking notes of what it has done, and transfers them (“this is what I did”) to the next agent. The next agent deals only with the notes.
You will recognize the architecture: this is Chapter 10’s hybrid pattern, third appearance, as promised. The next agent gets the metadata; the full transcript stays behind in referential storage, queryable if genuinely needed. This chapter is about designing the baton.
What a handoff note contains
A good note answers four questions. Here is the schema with a worked example:
## HANDOFF: Driller → Evaluator
WHAT HAPPENED: Completed objection drill, scenarios 1 and 3 (skipped 2 for time).
OBSERVATIONS: Strong on acknowledge-and-reframe; twice jumped to discounting
under price pressure; recovered well after hint in scenario 3.
USER STATE: Engaged, slightly fatigued after scenario 3; asked to see the rubric.
NEXT: Evaluate against rubric emphasizing the discounting pattern; user prefers
direct feedback (per memory).Four design rules govern every note:
Structured and small. A few hundred tokens. The note enters the next agent’s every-turn context, so Chapter 6’s token tax applies to it for the rest of that agent’s phase. A note that rambles is a tax that compounds.
Observations, not transcript quotes. The note carries the judgments the next phase needs (“twice jumped to discounting under pressure”), not the evidence it doesn’t. If the evaluator truly needs the exact exchange, it can search the referential store.
User state travels. Emotional context (fatigued, frustrated, energized) is exactly what naive handoffs lose and what human handoffs never would. A human trainer handing a student to a colleague says “she’s a bit worn down, go easy on the opener” without being asked. Your note schema has to ask.
A next-step directive. The note ends with what the receiving agent should do, resolving ambiguity at the seam. The receiving agent should never have to infer its own brief from raw observations.
Who writes the note, and when
The departing agent writes it, as its exit action. Concretely: the final step of each sub-agent’s session flow is literally “write the handoff note per the schema.” It is a flow step like any other, with Chapter 7’s rules applying.
And the schema lives in your design, not in each agent’s imagination. Give every sub-agent the same note template in its prompt. If each agent improvises what to pass forward, your seams will have four different shapes and the receiving agents will misread all of them. The schema above (what happened, observations, user state, next) is a solid default; extend it per domain, but keep it uniform across the session.
The transcripts, meanwhile, remain in referential storage, Chapter 12’s tier two. A receiving agent that genuinely needs detail can search for it, at the usual just-in-time latency. Metadata forward, detail on demand: the pattern holds at every seam.
Failure resilience
The property that made coding agents adopt notes in the first place: even if something fails, the next agent starts from the notes.
For conversation agents that buys three things:
- Graceful degradation. If the driller crashes or the connection drops mid-phase, the wrap-up agent still has the notes from every completed phase. The session degrades to “we lost the last exercise” instead of vaporizing entirely.
- A free audit trail. The notes are the session’s own summary of itself, phase by phase. When Part 7’s transcript mining begins, read the notes first: they localize problems to a phase before you open a single full transcript.
- Testable seams. You can verify handoffs deliberately: end a session mid-phase, then check that the next session’s greeter, reading memory built from partial notes, describes the partial progress correctly. Seams that survive that test survive production.
Notes handle the sequential case: phase A hands to phase B, in order. But who decides the order? And what happens when the right next phase depends on how the user is doing? That is a routing job, and jobs get agents. The supervisor is next.