Multi-agent orchestration

Chapter 21 · when one agent hits its ceiling

The hour

• Split roles: one job per agent
• Orchestration is control flow, not a framework
• Three patterns: pipeline, router, debate
• More agents = more cost + a new failure mode

Why more than one?

"Write it, critique it, rewrite it, fact-check, stay friendly" — the instructions fight each other.

A writer that only writes + a critic that only critiques beats one agent doing both.

Orchestration is just Python

def collaborate(task):
    draft    = agent(WRITER, task)
    critique = agent(CRITIC, f"DRAFT:\n{draft}")
    final    = agent(WRITER, f"Fix these:\n{critique}")
    return final

Passing outputs between agents is a function call.

Three patterns

• Pipeline — draft → critique → revise
• Router — classify, then dispatch to a specialist
• Debate — for / against / judge

Each is control flow over one-job prompts.

Router = an if over a classification

route = agent(ROUTER, message).strip()   # "billing"
answer = agent(SPECIALISTS[route], message)

Scale across domains without one bloated prompt.

The costs and the new failure mode

• 3 agents = 3× the calls, latency, and bill
• Agents can talk past each other

Pay when quality is worth it. Evaluate the OUTPUT, not your faith in the architecture (week 15).

Put it to work — three apps

• Writer + critic — the critique pass, made visible
• Router — classify → specialist
• Debate — for / against / judge

Takeaway

Split roles, compose with plain control flow. Pipeline, router, debate. Frameworks add convenience, not magic. Next: the capstone — the whole course in one app.