The shapes: supervisor–worker, peer handoff, debate

Lesson 1 of 5 in Multi-Agent Systems: When One Agent Isn’t Enough.

A multi-agent system is exactly what the workflow module warned you it would be: more than one agent loop, each with its own context window, coordinating on one task. That last clause is the whole subject. Two agents are not two employees; they are two separate context windows that cannot see each other’s memory and communicate only through the messages you route between them.

Three coordination shapes cover nearly everything shipping today. Learn them as communication topologies — who holds the goal, who holds the context, and where results flow — not as team metaphors.

Almost everyone arrives at multi-agent through the same door, so walk through it deliberately before we shut it. Picture one capable generalist who takes a piece of work from intake to delivery — and beside them a small team who split the same work: one gathers material, one drafts, one checks the numbers. The team’s appeal is real. Each member goes deeper on less, two of them can work at the same time, and nobody has to hold the entire problem in their head. That is a genuine pattern for humans, which is why an org chart is the first architecture almost everybody sketches on the whiteboard.

The problem is that the analogy arrives with luggage. A human team works because of things nobody writes down: colleagues refresh each other’s context constantly and for free, they argue with a brief that smells wrong, and somebody’s name is attached to the result. Remove those three and you keep the division of labour while losing everything that made it survivable.

What the team analogy quietly borrows — and what each borrowed assumption costs once the colleagues are agents
Borrowed assumption	Human team	Agent team	What it costs you
Sharing context	Ambient and nearly free. Someone leans over, asks, and ten seconds later two people know the same thing. Nobody itemises the exchange.	Nothing is ambient. Every fact one agent needs from another is re-transmitted explicitly as text, and both ends pay tokens for every word — context windows do not touch.	Coordination stops being invisible and becomes a measurable line item. Lesson 3 itemises the bill; what costs a human ten seconds costs you a brief, a round-trip, and a re-sent context.
Pushback on a bad brief	A specialist handed an incoherent brief stops and asks. Being publicly wrong is unpleasant, so people check first.	A worker agent executes the brief it was given — faithfully, competently, immediately. A brief that dropped a constraint still parses and raises no error.	The telephone game of lesson 4. The more capable the worker, the more thoroughly it finishes the wrong task.
Accountability	One name is on the outcome, and that person carries the lesson forward — the next brief is better because the last one hurt.	No component owns the mistake. Each agent did its job by its own instructions; the wrongness lives between them, and the trace is the only witness — if you kept it.	Root-causing shifts from asking a person to reconstructing several contexts, some of which the framework already discarded. Lesson 3 again.
Onboarding cost	Paid once. A new specialist absorbs the domain and the conventions, and that investment amortises over years of work.	Paid on every call. A worker is instantiated fresh, re-reads its system prompt and tool schemas, and forgets the entire engagement the moment it returns.	Specialisation is no longer an investment; it is a recurring per-call fee. A narrow role has to earn back its overhead on every single run.

Read the right-hand column as one sentence and the conclusion is uncomfortable: the shape of an org chart is not what makes a human team work. The chart is a communication structure that survives contact with reality only because the people inside it repair it continuously — backfilling context, renegotiating who owns what, refusing a brief that reads wrong. Copy the shape into software and you import the coordination cost in full while leaving the repair mechanism behind.

That is the trade this module prices out. The tax in lesson 3 is charged on day one whether or not your task had the shape you guessed; the failure modes in lesson 4 are simply the repairs that no longer happen by themselves.

What survives the puncture

Two things, and neither of them is hierarchy. Context isolation: a worker with a clean window can afford to read far more than the parent ever could, and hands back three sentences instead of fifty pages. Genuinely heterogeneous capability: a different model tier, a disjoint toolset, or a privilege boundary the runtime enforces — the agent that reads untrusted input physically lacking the credentials that would make it dangerous.

Both are properties of the context, not of the org chart — which is why lesson 2 states them in contexts, models and privileges rather than in job titles, and adds the one case where splitting genuinely refunds latency. Keep the analogy as the thing that got you here, and then stop using it as evidence: with the org chart gone, nothing argues for a second agent by default. One loop stays the starting point, and the burden of proof sits on the split.

So read the rest of this lesson as topology, not as staffing. Here are the three shapes themselves — for each one, ask the three questions that survived: who holds the goal, who holds the context, and where results flow.

Supervisor–worker

Task arrives
Supervisor decomposes task
One agent owns the goal end-to-end. It splits the task, writes a brief for each worker, and never does object-level work itself.
Worker A (fresh context)
Receives only its brief plus its own tools. Its context fills with search results and drafts the supervisor never sees.
Worker B (fresh context)
Runs in parallel with Worker A — the two share no state and cannot collide unless they touch the same external resource.
Supervisor integrates reports
Workers return compressed reports, not transcripts. The supervisor reasons over summaries — which is both the point and the risk.
Result

Supervisor–worker is a hierarchy: one agent owns the goal, subagents own subtasks, and control always returns to the top. It is the dominant shape in practice because it keeps one place where the whole task is understood — the same reason orchestration code was the star of the workflow module.

Peer handoff removes the hierarchy: agents pass ownership of the conversation sideways. A triage agent realizes the query is about refunds and hands off to the refunds agent, which now talks to the user directly. Nobody supervises; the handoff chain is the architecture. The handoff is a routing decision made by a model, so treat it with the same suspicion you treat any model-chosen branch.

Peer handoff

User query
Triage agent
A cheap generalist whose only real tool is choosing the right specialist.
Which specialist?
A model-made routing decision — log it, eval it, and give it an escape hatch to a human.
Billing agent takes over
Ownership transfers. The billing agent now holds the conversation, with its own tools and instructions.
Tech-support agent takes over
Each specialist can hand off again — chains longer than two or three hops are a smell.
Resolved

Debate / committee runs the same task through multiple agents and reconciles their answers — proposer versus critic, or N independent solvers plus a judge (often an LLM-as-judge). You are buying redundancy, not decomposition: nothing gets split, everything gets done twice or more. That makes debate the most expensive shape per unit of work and the easiest to justify only at high stakes — a security review before a merge, a medical draft before a clinician. The critic’s value comes from independence: a critic that shares the proposer’s context inherits the proposer’s blind spots.

The three shapes, side by side
Shape	Who owns the goal	What flows between agents	Native win	Native risk
Supervisor–worker	The supervisor, always	Briefs down, compressed reports up	Context isolation + parallel fan-out	Telephone-game briefs; supervisor reasons over lossy summaries
Peer handoff	Whoever holds the conversation right now	Ownership of the whole conversation	Specialists with narrow tools and narrow privileges	Routing errors; hot-potato loops between specialists
Debate / committee	Nobody — a judge picks or merges	Complete candidate answers + critiques	Catches single-agent blind spots at high stakes	N× cost for the same work; confident consensus on a shared wrong premise

Key terms: multi-agent system, supervisor–worker, subagent, handoff, orchestration, LLM-as-judge

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.