Multi-agent systems: when they help and when they're theatre
Splitting work across several agents adds latency, cost and failure modes. The cases where it earns them, and the cases where one agent is better.
Multi-agent diagrams are compelling: a researcher, a writer, a critic, all collaborating. In production, most of them would work better, faster and cheaper as one well-prompted agent with good tools.
What you pay
Every hand-off is a serialisation of context into text and back, which loses information. Every extra agent is another model call, more latency, more cost and another place for the chain to break. That price is real and it is charged per request.
Split into multiple agents, or not?
- Sub-tasks want different modelsSplit
- One component must not have write accessSplit
- Independent branches can run in parallelSplit
- A critic with a different objectiveSplit — this one reliably pays
- Same tools, same prompt, different nameKeep one agent
When it is worth paying
The cases that genuinely justify separation share one property: the sub-tasks want different things.
- Different models — a cheap fast one for classification, an expensive one for the hard reasoning
- Different tool permissions, where one component must not have write access
- Genuine parallelism, where several independent branches run at once
- Adversarial review, where a separate critic with a different prompt catches what the author misses
The critic pattern is the one that consistently pays
A second pass whose only job is to find fault with the first — prompted to look for specific failure modes, not to be generically helpful — reliably improves output quality. It works because it is genuinely a different task, not because it is a different agent.
What each hand-off costs
Real, and charged per request. Worth paying only when the sub-tasks genuinely want different things.
- Context lost in serialisation5
Everything becomes text and back again
- Added latency4
Another full model call in the chain
- Added cost4
Multiplied by every request
- New failure modes3
Another place the chain can break
Start with one
Build the single agent, find where it actually fails, and split only along that seam. Designing the org chart before you know the failure modes produces elaborate systems that fail in ways nobody anticipated.