multi-agent system

In depth

The one solid argument

The only solid argument in favour of separation is context. A single agent in charge of everything piles into its window the search results, the drafts, the errors and the retries, until the essential is drowned in the noise. Entrusting the search to one agent, the writing to another and the checking to a third lets each see only what concerns it, and pass on only a clean result. Then comes an argument from specialisation: different tools, instructions and models depending on the role, a small fast model to sort, a reasoning model to arbitrate.

The cost of coordination

The coordination cost is real and often greater than the gain. Every handover is an opportunity to lose information, every additional agent multiplies the calls and therefore the bill and the latency, and diagnosis becomes difficult: when the final result is bad, you have to work back up the chain to find who was at fault. Compound reliability applies between agents just as it does between steps, so that multiplying the participants degrades end to end success instead of improving it.

The practical rule

The practical rule is therefore conservative. You start with a single well equipped agent, you measure, and you separate only when a precise problem demands it: a context that overflows, a need for incompatible tools, a requirement to keep data apart. The most often justified pattern remains the simplest, that of a main agent delegating delimited tasks to subagents and collecting their results, without those subagents talking to one another. Architectures where several agents negotiate freely are spectacular in a demonstration and painful in production.