Multi-agent architecture is the fastest-spreading idea of the last two years. Splitting one agent into specialists sounds right, because human teams work that way too.
Two major reports from the field say the opposite of each other. The interesting part is that both are right at the same time.
Two opposing experiences
Anthropic described the multi-agent system it built for research work. A lead agent splits the job and sub-agents work on the pieces in parallel. In their internal evaluation the result came out 90.2 percent better than a single-agent setup (Anthropic).
Cognition wrote the exact opposite, and the title of the post was literally "don't build multi-agents". Sub-agents working in parallel make independent decisions, independent decisions on the same problem produce outputs that contradict each other, and context cannot be shared well enough between agents (Cognition).
Not a contradiction, different work
Put the two reports side by side and the distinction sharpens. What decides it is whether the work genuinely runs in parallel.
Research parallelises. Ten different sources can be scanned at once, the sub-jobs do not wait on each other's results, and the gathered pieces can be merged at the end.
Writing code does not parallelise. Every decision constrains the next one, and two agents deciding separately about the same file produces contradictions.
Anthropic's own post says as much. Domains that need shared context and have high dependency between agents are described as not yet suitable, and most coding work is said to contain far fewer parallelisable pieces than research does.
The gain has a measured price
The same post holds another number that does not get talked about much. Multi-agent systems use roughly fifteen times the tokens of a chat interaction. Even a single-agent setup uses four times as many as chat.
More striking still, in one evaluation 80 percent of the difference in performance was explained by token usage alone.
That finding does not make multi-agent architecture worthless, but it puts the expectation where it belongs. The system gives better results because it spends far more computation on the same problem. There is no free gain coming from the architecture itself.
Seen that way, the value of the work has to cover that cost. For work with expensive output, like legal review or competitor analysis, the sums work. For daily routine work they do not.
The real difficulty is sharing context
Cognition's most important warning is not about cost. They say sub-agents need to be given the full trace, not just the last message, because the details of how the work was split change how the task is interpreted.
There is a fine balance here and we wrestle with the same one. Giving a sub-task everything is expensive and scatters attention, while giving it too little produces work that looks reasonable and is wrong.
The distinction that worked for us was giving the sub-agent not only "what to do" but "why this work is being done". A sub-agent that knows the purpose makes a reasonable decision when it hits missing information. One that does not picks a route of its own.
We also choose deliberately what stays out of the briefing. Sub-tasks never see much of the main run's background, because that information does not change their decision and only takes up space.
Three questions before you split
Do the sub-jobs need each other's output? If they do, do not split. A single agent working in sequence gives a better result.
Can the results be merged? If the sub-agents' outputs can contradict each other, whatever resolves that contradiction is a cost item too.
Does the value of the work cover fifteen times the cost? Asking that question at the start is cheaper than asking it while looking at the invoice six months later.
The right order to start in
Start with one agent and do not split until it hits its limit. The limit usually makes itself known, because an agent trying to do very different kinds of work in one run starts to lose the context.
When you do split, the most natural seam is scanning and collecting work that does not depend on anything else. Reading five different sources at once splits well. Five people writing the same text together does not.
And do not build the structure between equals. One agent that leads and hands out work, with the ones doing the work beneath it, is what works. Equal agents talking freely to each other cannot pull off the coordination with today's models.
Sources
Anthropic, How we built our multi-agent research system
Cognition, Don't Build Multi-Agents