Parallelism

What this lesson demonstrates

This is the payoff of a graph. add_fan_out_edges broadcasts the same message to several workers that run concurrently; add_fan_in_edges gives a joiner a list of all their outputs, and the joiner fires only once every source has completed. Here one phrase fans out to a Shout worker and a Reverse worker, then a Join merges both — pure executors, zero credentials.

The code

Two edge calls express the whole parallel-then-merge shape:

workflow = (
    WorkflowBuilder(start_executor=dispatch)
    .add_fan_out_edges(dispatch, [shout, reverse])  # parallel branches
    .add_fan_in_edges([shout, reverse], join)       # synchronized merge
    .build()
)
result = await workflow.run("agent framework")
print(result.get_outputs()[0])

The fan-in target’s handler is typed to receive the aggregate:

async def run(self, results: list[str], ctx: WorkflowContext[Never, str]) -> None:
    await ctx.yield_output(" | ".join(sorted(results)))

What to notice

• Fan-out broadcasts the same message. Every worker in the list receives an identical copy of the source’s output — this is a broadcast, not a split.
• The joiner receives a list. Its handler signature is list[str], the aggregated outputs of all sources, and it runs exactly once.
• Fan-in is synchronized. The join only fires after every source completes, so wall-clock time is the slowest worker, not their sum.

The gotcha

Worker completion order is not guaranteed. The list handed to the joiner reflects concurrent completion, so its order can vary run to run. This lesson calls sorted(results) to get a stable output; if you need deterministic ordering, sort or key the results yourself rather than assuming input order.

How it maps to MAF and Foundry

Fan-out / fan-in is credential-free graph plumbing, but it’s exactly how you parallelize agent calls later: broadcast a task to several agents, then merge their answers in a joiner. Running it on pure string workers first makes the synchronization barrier obvious before latency and model cost enter the picture.

Run it

uv run tutorial/03-workflows/02_parallelism.py

Runs offline — no Azure creds. Success: one line containing both a REVERSE: and a SHOUT: segment from the two workers.

Next: Checkpointing