A team proposes five specialised agents that collaborate to handle a customer request. Assess.

What the interviewer is testing

Whether you can price the reliability cost of composing probabilistic components.

The concern

Error compounds multiplicatively. If each agent handles its step correctly 90% of the time, five in sequence gives 0.9^5 ≈ 59%. And unlike deterministic components, a failure is often not an error — it is a plausible wrong output that the next agent accepts as input and builds on.

Add the other costs: five times the inference cost and latency; non-deterministic control flow, so the same request takes different paths; and debugging that requires reconstructing a conversation between components rather than reading a stack trace.

The question to ask

What in this task genuinely requires agency rather than a defined sequence?

Most tasks framed as multi-agent problems are a known workflow: classify the request, retrieve relevant context, generate a response, validate it, route it. That is a pipeline with a model at each step, which is predictable in cost and latency, individually evaluable at each stage, debuggable, and testable.

The specialisation the proposal wants — different prompts and tools per step — is achieved without any of the coordination cost.

Where multi-agent genuinely helps

Open-ended tasks with no known decomposition, where the steps depend on what earlier steps discover. Genuine parallel exploration where several approaches are tried and the best selected. And tasks where an adversarial or critic role materially improves output — a generator and a reviewer is a well-supported pattern.

The recommendation

Start with the workflow. Define the steps, evaluate each independently, and measure the end-to-end result. Introduce agency only at the specific point where a fixed sequence demonstrably fails — supported by the evaluation, not by architecture preference.

What a strong answer adds

If multi-agent is genuinely warranted: bound the total steps and cost, validate the output of each agent before it becomes another's input, make the handoffs structured rather than free-text, and log the full trace. Most multi-agent failures are unvalidated handoffs.

Common weak answers

Approving because the framework supports it. Rejecting agents entirely without naming where they apply.