07 · Agent Teams & Orchestration

What you’ll learn

Agent teams add a layer of orchestration on top of multi-agent systems. A supervisor inspects each request, routes it to the most suitable team, lets that team run its own internal agents, then reviews and returns the final answer.

Try What is 144 / 12? (routed to the quant team) versus Summarise multi-agent systems (routed to the research team, which is the whole Project 06 pipeline running as a sub-team).

The orchestration pattern

flowchart TD
  Q[Request] --> S{Supervisor<br/>route}
  S -- numeric --> QT[Quant team<br/>calculator tool]
  S -- knowledge --> RT[Research team<br/>researcher → analyst → writer]
  QT --> S2[Supervisor<br/>review & merge]
  RT --> S2
  S2 --> A[Final answer]

import { agentTeamsPipeline, loadCorpus } from '@lib/js';

const corpus = await loadCorpus('knowledge-base', import.meta.env.BASE_URL);

agentTeamsPipeline('What is 144 / 12?', corpus);        // → quant-team
agentTeamsPipeline('Summarise multi-agent systems', corpus); // → research-team

// The trace shows the supervisor's `route` step and the delegated team's steps.

from _shared.data import load_corpus
from agent_teams import agent_teams_pipeline

corpus = load_corpus("knowledge-base")
print(agent_teams_pipeline("What is 144 / 12?", corpus)["answer"])
print(agent_teams_pipeline("Summarise multi-agent systems", corpus)["answer"])

Orchestration principles

• Routing. The supervisor’s only job is to send work to the right place — keep it simple and explicit.
• Encapsulation. Each team owns its internal agents; the supervisor sees only inputs and outputs.
• Review. The supervisor can validate or merge results before returning them.
• Scales down. When a task is trivial, a good orchestrator delegates to one agent rather than the whole org.

You’ve now gone from a single prompt to a coordinated organisation of agents. Revisit the Notes for the patterns, and the Resources for the papers behind each idea.