Multi-agent systems: when one agent isn't enough
A multi-agent system puts several AI agents on one task. A lead agent splits the work and hands pieces to worker agents, which run at the same time. The speed comes at a steep price — Anthropic measured roughly 15× the tokens of a chat — so prove one good agent isn't enough before you add more.
Using many agents sounds exciting. It is also the most expensive way to build. The real question is not whether several agents can share a task — it's whether the task is worth the cost of keeping them coordinated.
Two things push a task past a single agent. It may be too big for one context window — the agent's working memory. Or it may have parts that could run at the same time. The proven fix is orchestrator–worker. In plain words: one lead agent plans the job and splits it into pieces. Worker agents each take a piece and run in parallel, each with a fresh context window. The lead then combines their results. Anthropic's research system works this way — it beat a single-agent setup by 90.2% on their internal research test [1].
The same report shows the bill. A single agent uses about 4× the tokens of a normal chat. A multi-agent system uses about 15× — and token spend explained most of the difference in results. You are literally buying skill with tokens. That only makes sense when the task is valuable enough to cover it [1].
More agents means more working memory — and a bill for every one.
Why is coordination the hard part?
The wiring stopped being hard years ago — AutoGen showed general patterns for agents talking to each other back in 2023 [2]. What still goes wrong is the teamwork, not the tech: two workers doing the same job twice, details lost when work changes hands, and one worker's early mistake spreading to everyone after it. And if you already know how the work should be split, you don't need a lead agent at all — write the split into a plain workflow yourself, instead of paying an agent to figure it out each time [3].
The habit is the same as the whole build sequence: start with one agent. Measure where it actually fails — does it run out of memory? Is the task too wide? Add a second agent only when your tests show the gap.
Sources
- Anthropic — How we built our multi-agent research system, 13 Jun 2025
- Wu et al. — AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation, arXiv 2308.08155, Aug 2023
- Anthropic — Building effective agents, 19 Dec 2024 (orchestrator-workers as a workflow)
Frequently asked questions
What is a multi-agent system?
A setup where several AI agents share one task. A lead agent plans and splits the work. Worker agents each handle a piece at the same time, each with its own context window. The lead then combines the results into one answer.
When should I use multiple agents instead of one?
Only when you can measure the need: the task is too big for one context window, or it has parallel parts one agent handles too slowly. If you already know how to split the work, use a simple workflow instead — set the structure yourself and skip the coordination cost.
Why are multi-agent systems so expensive?
Every agent re-sends its whole growing history on each step, and a multi-agent system runs several agents plus a lead. Anthropic measured roughly 15× the tokens of a normal chat. You are buying capability directly with token spend.