Multi-agent patterns

Multi-agent patterns describe how several AI agents work together – instead of a single agent doing everything alone.

Why several agents instead of one?

A single agent has a limited context window and works step by step. For large tasks – extensive research, complex codebases – that runs into limits. Multiple specialized or parallel agents can get around this, but cost more tokens and need coordination.

Common patterns

Orchestrator-worker: a lead agent (orchestrator) breaks a task down, delegates parts to subagents, and combines their results. Anthropic's own research system uses exactly this pattern: a lead agent coordinates multiple specialized subagents that investigate different aspects of a question in parallel.
Sequential pipeline: the output of agent A becomes the input for agent B, which passes it on to agent C – like an assembly line.
Parallel fan-out: multiple agents work independently on different sub-problems at the same time, and the results are merged at the end.
Review/debate: one agent produces a solution, a second, independent agent checks or critiques it before it's approved.

The tradeoff

More agents means more tokens, more cost, and more coordination overhead – but also more parallelism, better specialization, and a cleaner main context, because intermediate steps stay in their respective subagent contexts instead of flooding the main context.