fan-out

flowchart TD
    Task([Task input]) --> W1[Worker agent 1]
    Task --> W2[Worker agent 2]
    Task --> W3[Worker agent N]
    W1 --> Aggregator[Aggregator agent]
    W2 --> Aggregator
    W3 --> Aggregator
    Aggregator --> Done([Merged output])

A task is split into independent subtasks, each handled by a separate agent in parallel. Results are merged by an aggregator agent.

How it works

The dispatcher breaks the task into N independent subtasks.
N worker agents run in parallel, each processing one subtask.
The aggregator receives all results and produces a single merged output.

When to use

Tasks that decompose naturally into independent units (e.g. reviewing multiple files, generating multiple variants, analyzing multiple dimensions).
Situations where parallelism meaningfully reduces wall-clock time.

When not to use

Tasks where subtasks depend on each other — use a sequential pipeline instead.
When N is unbounded or large enough to hit rate limits or incur significant cost.

Trade-offs


Pro	Wall-clock time scales with the slowest subtask, not the sum
Pro	Each worker is focused and isolated — easy to reason about
Con	All subtasks must complete before the aggregator can run
Con	A single slow or failed worker blocks the merge step

Failure modes

Partial failure — one worker fails; the aggregator receives incomplete results and may produce a misleading summary.
Cost explosion — N subtasks × M tokens each; unbounded fan-out is expensive.