The correct answer is the **market-research workflow**. The other three are single-shot tasks that gain nothing from multi-agent and pay coordination overhead.
**Why market research wins.** Eight independent sub-questions means eight researcher agents can run in parallel, each handling one sub-question. Wall-clock time drops from 'eight retrievals serialised' to 'roughly one retrieval' because they run at the same time. A final reducer agent reads all eight findings and synthesises the report. The total token spend is similar to the single-agent version, but the user-perceived latency is dramatically better.
**Why the other options lose.**
- **Single customer-support question:** one retrieval, one reply. No sub-tasks to fan out across. Multi-agent adds an orchestration step for zero benefit.
- **Single tweet from a prompt:** one short generation. There is nothing to parallelise. Splitting into agents would mean two prompts pretending to be a workflow.
- **Email classification:** one short classification call. Multi-agent would add overhead and possibly disagreement between classifiers without improving accuracy on a simple decision.
**The decisive properties of a real fan-out workflow** are independence (the sub-tasks do not need each other's outputs) and a clean join (the synthesis step can combine the results into the final answer). Both have to be present. Without independence, parallelism does not work (one sub-task waits on another). Without a clean join, the parallel results pile up with no way to combine them.
**Parallel fan-out** is the canonical multi-agent win when the workflow has independent sub-tasks plus a clean join step. The market-research option is the textbook fit; the other three are single-shot or single-tool tasks where multi-agent is overhead.
**The market-research case.** Eight sub-questions about a market overview (competitors, pricing, market size, growth rate, key players, recent moves, regulatory environment, customer segments). Each sub-question is independent: the answer to one does not depend on the answer to another. Each can be handled by a researcher agent that runs its own retrievals and produces a short brief. The eight agents run in parallel; wall-clock time is the latency of the slowest single sub-question rather than the sum of all eight. A reducer agent reads the eight briefs and synthesises the final report.
This is one of the cleanest multi-agent patterns. Total token spend is similar to a serial single agent (eight retrievals happen either way), but the latency improvement is 6-8x for the user. For research-style workflows where the user is waiting, this matters enormously.
**Why the other three fail.**
- **Single customer-support question:** the workflow is retrieve once reply once. No independent sub-tasks. Splitting into agents would add an orchestration round-trip for negative benefit.
- **Single tweet draft:** one short generation. No fan-out opportunity. A second agent would have to either duplicate the work or critique the first, which is a critic-loop pattern, not fan-out.
- **Email classification:** one decision over three labels. Multi-agent ensemble could be argued for accuracy boost, but it is best-of-N voting, not fan-out, and is overkill for a 3-class classifier where a small fine-tuned model nails it cheaper.
**Structural test for fan-out:** (1) Can I list the sub-tasks and verify they do not depend on each other? (2) Is there a clean way to combine the outputs? Both yes means fan-out works.
Quick reference
TL;DR
Parallel fan-out wins on workflows with multiple independent sub-tasks plus a clean join. Single-shot or single-tool tasks gain nothing from multi-agent and pay overhead.
Picture eight friends helping you research what restaurant to pick: one checks reviews, one checks the menu, one checks prices, one checks how busy it gets, and so on. They can all look at the same time, then meet up and tell you what they found. That is much faster than asking one friend to do all eight checks one after another. But if the question is 'what is two plus two', asking eight friends does not help, there is no work to split. Fan-out is for jobs with lots of independent pieces, not for one-step questions.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Single-shot tasks like classification, short generation, or single-retrieval Q&A There are no parallel sub-tasks to fan out across; orchestration overhead is pure cost.
Workflows where sub-tasks depend on each other's outputs Dependencies force serialisation on the dependent steps; the parallel guarantee breaks.
Batch workflows where latency does not matter The fan-out value proposition is latency reduction; if no user is waiting, the orchestration overhead is unjustified.
Quick scan · 2–4 min
Everything important, quickly.
6 to 8 min: the three structural requirements for fan-out, the latency vs cost framing, how to distinguish fan-out from best-of-N and critic patterns, why each wrong option fails the structural test, and production framework implementations.
Real-world examples
Real products, models, and research that use this idea.
1Perplexity's 2026 research mode runs parallel sub-question agents and synthesises into a cited answer in roughly the time of a single retrieval.
2Google's Gemini Deep Research feature fans out sub-queries across a research plan, then reduces into a report, exactly matching the canonical pattern.
3OpenAI's o1-pro and o3 reasoning models internally fan out exploration paths in parallel and reduce via a verifier step.
4LangGraph's Send API in 2026 production deployments commonly parallelises retrieval across 5-10 sub-queries for research-shaped workloads.
5Anthropic's Claude Projects with multi-document research surfaces parallel evidence gathering steps that fit the same shape.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you implement the market-research fan-out in LangGraph?
Sign in or go Pro to see the approach for this follow-up.
QWhat happens if one of the eight parallel researcher agents fails or times out?
Sign in or go Pro to see the approach for this follow-up.
QWhere does best-of-N fit in this picture, and how is it structurally different from fan-out?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Picking 'one customer-support question with a retrieval' because the word 'retrieval' sounds agent-flavoured. A single retrieval is a single tool call; there are no parallel sub-tasks to fan out across.