We expected a team of cooperating agents to beat one. It mostly added coordination problems.
A thought to take with youDefault to one well-equipped agent with clear tools. Split into several only when there is a real reason.
What we expected
Splitting a task between several specialised agents, each with its own role, sounded like a natural way to handle complex work.
What happened
In our prototypes, a single agent with good tools was usually faster and more accurate. The multi-agent versions spent effort passing messages back and forth, sometimes disagreed about the state of the work, and repeated each other’s mistakes. They were also harder to debug.
What we changed
We now start with one agent inside a simple, predictable workflow. We split the work only when there is a security boundary to respect or tasks that genuinely benefit from running in parallel.