Go back to Chapter 7’s demo prompt, the five-phase session flow: greet and set context, present the situation, check understanding, go deeper, wrap up. You have seen that exact flow as sections of one prompt.
Now watch it become separate agents, because the mapping is one-to-one: phase headers become agent boundaries. The numbered flow you already know how to write is the decomposition diagram. That is the deep convenience of cutting along phase boundaries rather than topics: each phase already has its own job, its own tools, and its own exit condition, which is precisely what an agent needs to be.
The standard decomposition
Most scenarios decompose into the same trio. Learn it as the default, and vary from there.
Agent 1: the greeter (30 to 60 seconds). Job in one line: welcome the user, establish continuity, set today’s agenda. The context it carries is minimal by design: the persona block, the super memory (Chapter 12), and today’s agenda template. It knows who it’s meeting, it says hello, it sets the context, and that is all. Tools: usually none. Exit condition: agenda acknowledged. Write the handoff note, transfer.
Agent 2: the core (the bulk of the session). Job: the actual work (teach, drill, roleplay, interview). Context: its phase instructions, the greeter’s note, and the scenario content (or its knowledge-base catalog lines, per Chapter 10). Tools: the full working set (image generation, MCQ, knowledge search).
For complex sessions the core itself splits further: a teacher (present the concept, use slides and images) followed by a driller (run the practice, apply Chapter 18’s pressure), each with its own exit. Exit condition for any core agent: phase objectives met, or time budget spent.
Agent 3: the wrap-up (1 to 2 minutes). Job: summarize, deliver feedback, preview next session, write memory. Context: the notes from all prior phases, not their transcripts, a distinction Chapter 16 makes precise. Tools: memory write and end-session. Exit: the session ends deliberately, using Chapter 20’s closing pattern.
Scoping each sub-agent’s prompt
Each sub-agent is a full Chapter 7 citizen, in miniature: all four blocks, scoped to its phase:
- Role and character: the shared persona core plus a phase-specific stance: “you are in evaluation mode: measured, specific, kind.”
- Flow: the phase’s internal steps with time budgets, exactly like a full prompt, just shorter.
- Resistance and pressure: only where they belong. The driller pushes back; the greeter never does; the evaluator is honest but kind. Splitting lets each phase have exactly the friction it needs, rather than one blended temperament.
- Principles: the phase’s do’s and don’ts, within the 10 to 15 negative-rule cap per agent. A quiet bonus of splitting: each agent gets its own budget, so the constraint that pinched your monolith relaxes.
And the persona consistency rule, which prevents the felt-handoff problem from Chapter 14: extract a shared persona block (name, voice, language, warmth) and paste it into every sub-agent’s prompt verbatim. The user must experience one coach who changes activities, not three employees who share a headset.
Choosing the cuts: a procedure
- Write your single-agent session flow. You have it already from Chapter 7: numbered phases with time budgets.
- For each phase, list three things: the instructions it uses, the tools it uses, the context it needs.
- Cluster phases with high overlap: they stay in one agent. Cut where the overlap drops: where two adjacent phases share few instructions and fewer tools, that seam wants to be an agent boundary.
- Write each candidate agent’s job in one line. If you can’t, the cut is wrong: merge it back or re-cut. The one-line test is the cheapest validation in this playbook and it catches almost every bad decomposition.
- Define each seam: the exit condition, and what the handoff note must contain, which is the entire subject of the next chapter.
Example: the 30-minute PM interview coach, decomposed. Greeter (super memory, agenda). Case interviewer (case content, pressure instructions, image tool for exhibits). Behavioral interviewer (question-bank catalog lines, follow-up rules). Evaluator (rubric, MCQ for self-assessment). Wrap-up (feedback, memory write). Five agents, each with a one-line job, each prompt under 2,000 tokens, versus the 12,000-token monolith it replaced. Every agent operates on Chapter 5’s sharp end of the accuracy curve.
Five agents means four seams, and everything now depends on what crosses them. The baton matters more than the runners: state transfer is next.