Return to the 30-minute training session. But this time the user is ahead of schedule. They ace the quiz in two minutes flat.
A fixed linear pipeline drags them through the planned remedial drill anyway, because the pipeline knows only its sequence. A supervised session notices (the quiz agent’s handoff note says “scored 5/5, visibly under-challenged”) and routes to the advanced roleplay instead, skipping a phase entirely.
That routing decision is a job. And in this architecture, jobs get agents. The supervisor is the agent whose job is deciding which agent runs next.
When you need one, and when you don’t
Be suspicious of supervisors before you are enthusiastic about them, because most sessions don’t need one.
Linear flows: no supervisor. Greeter, then core, then wrap-up, in fixed order, needs only exit conditions and handoff notes. The “orchestration” is a hardcoded sequence, and hardcoded is a feature: fewer moving parts, nothing to misroute. Most first products live here, and should.
Branching flows: supervisor. The moment the next phase depends on what just happened, routing becomes a real decision. Route by performance: remediate or advance. By user choice: “practice negotiation or conflict today?” By time remaining: skip to wrap-up when five minutes are left, regardless of phase. By detected state: frustration shortens the session, energy extends it.
Portfolio flows: supervisor. A course with many scenarios, where each session should pick the right next module based on accumulated memory. The supervisor reads memory, assembles today’s plan, and dispatches: a curriculum director, not just a traffic cop.
What the supervisor holds
The supervisor is itself a context-engineering exercise, and the discipline of this whole playbook applies to it: minimal, sufficient.
It holds three things:
- The session plan: the phases available, their objectives, their time budgets.
- The routing rules. Declarative, in its prompt: “If quiz score is below 3, route to remediation. If time remaining is under 5 minutes, route to wrap-up regardless of phase.”
- The notes so far: its only view of what happened. The supervisor never reads transcripts; it reads batons.
And note what it does not hold: scenario content, rubrics, persona details. Those belong to the workers. A supervisor that starts accumulating worker context is drifting back toward the monolith you escaped in Chapter 14: the same disease, one level up.
One design choice to make explicitly: does the supervisor ever speak to the user? Usually it is silent: routing is invisible, and the sub-agents share the persona block so transitions feel like one coach changing activities. Occasionally it is voiced as a light narrator: “Great, let’s move to practice.” Either works. Mixing them (sometimes silent, sometimes speaking) feels haunted. Pick one.
On Tough Tongue: this is the super agent configuration. You define a flow: a supervisor agent, and beneath it the different agents that do the different pieces. You author each sub-scenario and the flow between them; the platform runs the supervision loop, the handoffs, and the note-passing. The pattern in this chapter is exactly what you are configuring when you build one.
The complete picture
Assemble everything Part 4 has built, because it is worth seeing whole.
The user talks to one persona-consistent voice. Behind it, the active sub-agent (greeter, teacher, driller, evaluator, or wrap-up), each carrying a small scoped prompt, its phase’s tools, and the shared handoff-note schema. Above them, the supervisor with the plan, the routing rules, and the notes. Beside them, the shared substrate from Parts 2 and 3: the knowledge base consulted just in time, two-tier memory, and guardrails around everything.
Every box in that picture is something you have now built, chapter by chapter. Which supports the claim this part opened with: sub-agent architecture is not a new discipline. It is Chapters 5 through 13 applied at a finer grain: smaller workspaces, engineered with the same rules, connected by notes.
The architecture is complete. What remains is making the conversations inside it feel real: the resistance, the guardrails, and the session flows that open and close like a human’s. That is Part 5.