Multi-agent system
What is Multi-agent system?
A system where several AI agents, each with its own role or tools, work together on a task too big or too varied for one. A lead agent often plans and delegates while specialised agents run parts in parallel. It buys breadth and speed, at the cost of more coordination, more tokens and harder debugging when something goes wrong.
Why it matters
The honest question is when the overhead is worth paying, because for most tasks it is not. Splitting work across several agents buys something only when the task is genuinely broad or parallel, with parts that can run independently and be recombined. Wrap a simple job in that machinery and you inherit the costs without the benefit: agents that duplicate each other’s work, errors that compound as one builds on another’s mistake, loops that burn tokens going nowhere, and a debugging problem that worsens with every participant, since a failure can now hide in the handoffs between agents rather than inside any one of them. The pattern is worth its cost on the class of problem no single agent can hold at once, and wastes it everywhere else.
In practice
A broad research task, gather from many sources then synthesise, splits naturally. Sub-agents chase different threads in parallel while a lead agent plans and stitches the findings together, and the whole thing finishes faster than one agent working in sequence. The same architecture around a single, linear task just adds coordination, cost and more places to fail, for a result one agent would have reached alone.
