
Multi-agent system: several AI agents splitting a job
In short: A multi-agent system splits one task across several AI agents with separate contexts, buying parallel work at a much higher token cost.
A multi-agent system is a setup where several AI agents, each with its own instructions, tools and context, work on parts of one task and exchange results. Common shapes are a lead agent delegating to workers, a pipeline of specialists, or agents critiquing each other. It buys parallelism and cleaner contexts at a price: Anthropic reported in June 2025 that such systems use about 15 times the tokens of chat.
Multi-agent system in plain words
One person writing a report versus a small team: one researches, one drafts, one edits. The team is faster on big jobs, and slower and noisier on small ones because everyone has to be briefed.
How it works
A coordinator breaks the goal into subtasks and starts worker agents with focused prompts. Each worker runs its own loop with a fresh context window and returns a compact result. The coordinator merges results, resolves conflicts and decides whether more work is needed. Anthropic's December 2024 guide calls this the orchestrator-workers pattern.
Anthropic's research system works this way: a lead agent plans, spawns subagents that search in parallel and synthesises what they return. On its internal research eval, a Claude Opus 4 lead with Claude Sonnet 4 subagents outperformed a single Claude Opus 4 agent by 90.2%.
Example
As an illustration, a market scan: a lead agent spawns five workers, one per competitor, each searching and summarising in parallel. The lead compares the five summaries and writes the brief. Each worker keeps a small, clean context.
Where you'll meet it
- Deep-research products that fan out searches
- Coding agents that spawn subagents for exploration or review
- Agent frameworks and SDKs with orchestrator and worker roles
What it changes for you
Token spend rises fast. Anthropic's write-up puts single agents at about 4 times the tokens of a chat and multi-agent systems at about 15 times, so the economics only work on tasks that are broad, parallel and valuable. It calls them a poor fit where agents must share one context or depend heavily on each other, which it says covers most coding tasks. Debugging gets harder too: you need traces that show which agent said what.
Multi-agent system vs subagent
A subagent is one helper spawned by a main agent for a bounded task. A multi-agent system is the overall design where several agents coordinate as peers or a hierarchy.
Common mistake
My view: splitting into agents by job title ("planner", "coder", "tester") because it looks tidy on a diagram. Split by context instead: give a separate agent a task when it needs a separate, smaller context.
Related terms
Subagent · Orchestrator-worker pattern · AI agent · LLM tracing · Cost per task
From the archive: Agent swarms need aggregation evals
Give a task its own agent only when it needs its own smaller context.


