Parallel Fan-Out

When to reach for this

Reach for parallel-fan-out when independent perspectives are needed simultaneously:

• Multiple aspects of the SAME input, no causal dependency. E.g. "sentiment AND topic AND PII detection" — none of these need the other's output.
• Latency matters. Running N LLM calls in parallel ≈ 1 LLM call's latency, vs. N× if sequential.
• Results are joined back together. A follow-up code block typically merges the fan-out outputs into one structured result.
• The branches are NOT competing. Use parallel-fan-out for "and" cases, not "or" cases. (If you're picking the best of N answers, you want a different shape — usually a quality-checking second block.)

If any branch depends on another branch's output, this is sequential-decomposition wearing the wrong hat.

Multiple independent analyses against the same input, joined by a final aggregator.

Flow: "Multi-Lens Content Review"
├── Container (parallel)
│     ├── Block: "Tone Check" (llm)
│     ├── Block: "Readability Score" (llm)
│     └── Block: "Fact Check" (llm)
└── Block: "Combine Verdicts" (llm)
      Input: { tone_check, readability_score, fact_check }

When to use: the analyses don't depend on each other. Parallel execution gives you wall-clock speedup roughly equal to the slowest branch, instead of the sum.

Common mistake: using a sequential chain when the steps are independent. If block B doesn't read block A's output, putting them in sequence is pure latency tax.