Multi-Agent Systems

You stepped through one agent loop. A multi-agent system connects several of those loops, passing tasks and evidence between them until a shared goal is complete.

Common coordination patterns include:

• Manager and workers — one agent divides the goal and delegates pieces
• Pipeline — output moves from researcher to writer to reviewer
• Debate — agents propose different answers, then a judge compares evidence
• Shared board — agents claim tasks and post results in one workspace

Each agent needs a role, allowed tools, inputs, an output format, and a stopping rule. The handoff matters as much as the agents: “research this” is vague; “return three claims with source links” is testable.

More agents do not create more intelligence for free. They add model calls, latency, duplicated work, and new failure points. A single well-designed agent is often the better system.

• Research where independent source-finding and fact-checking help
• Software tasks with separate implementation and review stages
• Large workflows that divide cleanly into specialist roles
• Experiments where competing approaches are useful

Agents can agree on the same wrong assumption, endlessly hand work back, or overwrite each other. Limit rounds, keep an audit trail, and define who makes the final decision.

Do not use extra agents as a substitute for good tools, clear instructions, or human approval on risky actions.

Use the Try yourself board above. Give one agent the planner role and another the checker role, then make each handoff include a concrete deliverable.

Take a familiar task and ask: could two independent roles improve it, or would they only add ceremony?