Do small specialised agents coordinating beat one frontier agent at real engineering work?

The intuition says yes: a swarm of small agents, each with a narrow tool and clear contract, should be cheaper, faster and more auditable than one frontier model holding everything in context. Real engineering work is less sure.

Coordination has a cost, and it is paid in the hardest currency: shared understanding. Specialised agents are only as good as the interfaces between them, and the interfaces are where ambiguity — the thing agents handle worst — accumulates.

The honest experiment is a bake-off on real tasks: same codebase, same goal, one frontier agent versus a coordinated set of specialists, scored on correctness and total cost. Until that’s run, this stays a question rather than a position.