Multiple Agents

Two agents are not twice as good as one. They are one system with a communication problem. Used well, splitting work across specialists raises quality because each one has a narrow job and a short prompt. Used badly, you have doubled your cost and added a new way to fail.

The five patterns worth knowing

Pipeline

A to B to C, each stage improving the last. Drafter, critic, editor. Simple, predictable, easy to debug. Start here.

Supervisor and workers

One agent owns the goal and hands tasks to specialists. The supervisor is the only one who knows the whole picture.

Critic

A second agent whose only job is to attack the first one's output. Cheap, and it catches more than another round of the same prompt.

Debate

Two agents argue opposite positions, a judge decides. Good for choices with real tradeoffs, expensive for everything else.

Blackboard

Agents read and write one shared state instead of messaging each other. Scales past the point where message passing gets tangled.

The rule of thumb

Add an agent when you can name its job in four words and say what it alone is accountable for. Otherwise add a tool, not an agent.

Optional: load a real model

A live pipeline: Drafter, Critic, Editor

Paste a rough note. The Drafter turns it into a message, the Critic attacks that draft against fixed checks, and the Editor applies the criticism. Each agent sees only what the one before it produced, which is what makes this a pipeline and not one long prompt.

1. Drafter

waiting

2. Critic

waiting

3. Editor

waiting

Want five agents and real tool calls?

The Multi-Agent Lab runs a Planner, a Librarian that searches the real course catalog, a Curriculum Architect, a Quiz Master, and a Reviewer that verifies every course the others cited and repairs invented titles. It is the fullest example on this site of agents handing work to each other and checking each other's output.

Open the 5 agent lab
The failure nobody expects. In a chain of agents, errors compound quietly. If each stage is 90% reliable, three stages give you about 73%, and nobody stage looks broken in testing. That is why the last agent in a serious pipeline is almost always a checker with deterministic rules rather than another writer. Next: routing work to the right specialist.

Free from AI School - no signup, everything runs in your browser.