Multi-agent systems: coordination is the actual hard part

Multi-agent architectures are seductive because they map onto how humans organize work: specialists, a coordinator, a hand-off protocol. The first time you split a complex task between a “researcher” agent and a “writer” agent, the result is genuinely better. The third time, you’re debugging a coordination bug that single-agent versions never had.

Where the coordination tax appears

Agents disagree about what’s already been done. Agents step on each other’s work because the shared state is updated asynchronously and read inconsistently. Agents loop because two of them keep delegating the same step back and forth. The coordinator agent — which everyone introduces around month two — solves some of this and adds its own failure mode: a single point of routing logic that the model has to get right every time.

When the cost is worth it

For tasks with genuinely heterogeneous subtasks — say, a creative writing step plus a fact-checking step plus a formatting step — specialization pays for the coordination overhead. For tasks that are uniform but long, a single-agent loop usually wins. The honest test: would a human solo operator do this task differently than a small team? If the answer is “no,” the multi-agent version is probably overhead pretending to be architecture.

Multi-agent systems are not faster, simpler, or more reliable than single-agent loops by default. They are sometimes worth the coordination cost — and the teams that ship them have learned to be skeptical about when.