ArchitectureIntermediate

Router vs supervisor

“What is the difference between a router architecture and a supervisor architecture, and when would you use each?”

What this tests

  • Precise understanding of two commonly confused patterns
  • Awareness of latency, cost, and failure differences
  • Ability to pick based on whether the task needs one path or a coordinated sequence
  • Whether they treat the supervisor as a risk and a bottleneck

Answers by level

Read the beginner answer first and notice what is missing.

A Router Architecture makes one decision at the start: classify the input and dispatch it to one of several specialised handlers, each of which completes the task on its own. The router is usually a cheap classifier or a small model with structured output; it runs once and exits the picture. A Supervisor Architecture stays in the loop: it decomposes the task, delegates sub-tasks to workers, receives their results, and decides what to do next, possibly many times.

Use a router when the inputs fall into distinct categories that need different handling and each category can be handled end-to-end by one path: support intents, document types, languages. Use a supervisor when a single task genuinely requires coordinating several capabilities whose results feed each other, such as research → analysis → drafting where the analysis result changes what to research next.

The supervisor is much more expensive: every round trip is an LLM call, context grows with worker outputs, and the supervisor becomes a single point of failure. Routers are cheap, fast, and testable as classifiers.

Green flags · Red flags

Green flags
  • Router decides once; supervisor stays in the loop
  • Router treated as a classifier with measurable accuracy
  • Mentions pipeline as the cheaper alternative to supervisor
  • Names supervisor cost, latency, and single point of failure
  • Handles ambiguity with an "unsure" class or clarifying question
Red flags
  • Calls the supervisor a "better router"
  • No cost or latency comparison
  • Defaults to supervisor for any multi-capability task
  • No plan for misroutes

Follow-up questions

F1
How do you evaluate a router?
F2
Supervisor keeps delegating the same sub-task. What do you do?

Practical scenario

A document-processing product receives invoices, contracts, and resumes, each needing different extraction. A colleague proposes a supervisor agent that "figures out what to do" for each file. Propose an alternative, explain how you would measure both, and describe under what evidence you would accept the supervisor.

Related concepts · Learn this topic