Single Agent vs Multi-Agent
A single agent handles a task from start to finish on its own. A multi-agent system splits the task across several agents working at once, then combines what they find. The real tradeoff isn't capability — it's cost: by one real accounting, a multi-agent system runs at roughly fifteen times the token cost — the unit both usage and price are measured in — of a single normal reply. Multi-agent earns that cost only when a task genuinely breaks into independent pieces worth running in parallel; otherwise a single agent does the same job for a fraction of the price.
What a single agent does
One AI agent works through the whole task itself — deciding, acting, observing the result, and deciding again — in one shared context. Every step sees everything the earlier steps found, which is exactly what makes it the simpler, cheaper option whenever the task doesn't actually need to be split up.
What a multi-agent system does
Several agents work on separate pieces of the task, often in parallel, each in its own separate context instead of one shared pool. A lead agent typically coordinates: breaking the task apart, handing pieces to worker agents, and combining what comes back. The Multi-Agent System page covers the mechanism and a real case study — Anthropic's own research system solving a many-company lookup task that a single agent couldn't complete in reasonable time.
Side by side
| Single agent | Multi-agent | |
|---|---|---|
| Best for | Tasks where each step depends on the last one | Tasks that genuinely split into independent pieces |
| Speed on parallelizable work | Sequential — works through pieces one at a time | Faster — pieces run at the same time |
| Context | One shared context for the whole task | Each agent keeps its own, separate from the others |
| Relative cost | Baseline | Roughly 15x a single normal reply, by one real provider's own numbers |
| Coordination needed | None — one agent, one path | Real — someone has to combine and reconcile what comes back |
| Failure handling | One agent's mistake is the whole failure | One agent's mistake needs to be caught by whatever combines the results |
Which one your problem calls for
Pick a single agent whenever it could plausibly finish the task — most tasks. If the steps depend on each other in sequence, or the task is small enough that parallel work wouldn't meaningfully speed anything up, splitting it across agents adds coordination cost for no real benefit.
Pick multi-agent when the task decomposes into genuinely independent pieces and finishing faster or covering more ground is worth paying roughly fifteen times the cost for. Looking up one fact about each of five hundred companies is the textbook case — each lookup is independent, so running them in parallel across several agents finishes faster with no piece waiting on another's result. Investigating a single, sequential bug — where step two only makes sense once step one's finding is known — is the opposite case: there's nothing to parallelize, and a single agent working through it in order is both cheaper and simpler.
A middle ground
The choice isn't always all-or-nothing. Some systems run as a single agent by default and only fan out into a multi-agent pattern for the specific sub-task that's genuinely parallelizable, folding the results back into the single agent's context once that piece is done. That keeps the expensive multi-agent cost scoped to the part of the task that actually benefits from it.
In this guide
FAQ
Does a multi-agent system require several different models, or different providers?
No — it's usually the same model running as several separate instances, each with its own context, not a mix of different models. What makes it a multi-agent system is the separation into independent working contexts and the coordination between them, not which model or how many different ones are involved.