Multi-Agent Coordination

“Multi-agentAgentA worker you delegate to: brief it, and it takes steps on its own. A chatbot answers; an agent acts.An LLM that runs tools in a loop toward a goal — it acts, checks the result, decides the next step, and repeats until done.Full definitionClose” is not one architecture. The decisive variable is coordination topology: who can create work, who owns shared state, how results move, and whether workers communicate only through a coordinator or directly with peers.

Three Topologies

1. Coordinator and workers — a lead decomposes work, spawns isolated agents, and collects their results.
2. Peer team — agents have independent sessions plus direct messaging and shared task state.
3. External orchestrator — a workflow engine owns scheduling, retries, state, and evidence while agents perform bounded steps.

Claude Code's documented agent teams use independent teammate sessions, a shared task list, mailboxes, and peer messaging, but Anthropic still labels the feature experimental. Codex documents parallel subagents whose results return through the main coordinator; the reviewed source does not describe a peer mailbox or shared task ledger. Copilot CLI Fleet provides coordinator-led parallel execution and documents sibling-agent messaging.

Those are semantic differences, not naming differences.

When Teams Pay

Use multiple agents for genuinely independent work, competing hypotheses, or separate review perspectives. Avoid them when tasks share files, depend on the previous result, or cannot state a crisp completion contract.

The cost model includes more than tokens: coordination messages, duplicate context, merge conflicts, verification, and recovery from partial failure. Start with the smallest useful fan-out and measure elapsed time, defect rate, and total model usage against a single-agent baseline.

The deeper lenses retain Claude Code team examples. Do not translate a peer-team design directly onto a coordinator-only surface without replacing its communication and shared-state assumptions.