StoryboardAI Engineering

Teams of agents

A lead agent hands parts of a job to helper agents, each with its own fresh window.

What is a team of agents?

One agent can split a job and give each part to a helper. Each helper is a separate agent with an empty window that knows only the brief it was handed. It does its part and sends back a short report.

Two things are gained. The parts run at the same time, so the job finishes sooner. And all the reading each helper does stays in its own window, so the lead's window holds three short reports instead of fifty web pages.

Two things are paid. A team uses several times more tokens than one agent. And every hand-off loses detail, because a helper knows nothing the brief did not say. Teams suit jobs whose parts are independent; for small or tangled jobs, one agent is better.

One lead, several helpers with their own windows.Step 1: A lead agent splits the job. Step 2: Helpers work side by side. Step 3: Each returns a short result. Step 4: The lead combines them.1A lead agentsplits the job2Helpers work sideby side3Each returns ashort result4The lead combinesthem
One lead, several helpers with their own windows.

Why does a team of agents matter?

One agent has one window. A big job fills it with material that is only needed briefly, and some work could run side by side if there were more than one worker. Splitting a job between several agents keeps each window small and gets independent parts done at the same time.

How does a team of agents work?

  1. A lead agent breaks the job into parts that do not depend on each other.
  2. For each part it starts a helper: a new conversation with its own empty window, a narrow role and a few tools.
  3. It gives each helper a complete written brief.
  4. The helpers work at the same time and each returns a short result.
  5. The lead agent combines the results. Only those results enter its own window.

When should you use a team of agents?

Use it when

  • Parts of a job are independent
  • A side task needs a lot of reading and yields little
  • A fresh pair of eyes should check the work

Think twice when

  • The parts depend on each other step by step
  • The job is small. A team costs several times more tokens

Where does a team of agents sit in a real system?

Starting a helper is itself a tool the lead agent can ask for. The harness creates each helper as a separate loop with its own limits, and returns its final message as a tool result.

What goes wrong with a team of agents?

  • Writing a thin brief. A helper knows nothing of the conversation and will guess.
  • Letting two helpers edit the same files at once.
  • Using a team for a task one agent could finish quickly.

What does a team of agents look like in practice?

A research question is split into wind tests, owner reports and prices. Three helpers each read about ten pages and return nine notes. The job takes 4 minutes, not 12, and uses 640,000 tokens where one agent would have used 180,000.

Watch it happen

These stories show teams of agents at work, one step at a time.