Some tasks don’t need a sequence — they need several pairs of eyes at once. Parallelization fires multiple LLM calls simultaneously, then merges the results. It comes in two flavors: sectioning (divide the work) and voting (repeat the work and compare).

The analogy

  • Sectioning is a team of proofreaders, each taking one chapter: the book is checked in a tenth of the time.
  • Voting is a medical second opinion: three doctors examine the same scan independently; if two see a problem, you take it seriously.

The principle

Sectioning — split the work:

flowchart LR
    IN([input]) --> A["LLM: part A"] & B["LLM: part B"] & C["LLM: part C"] --> M[merge] --> OUT([output])

Voting — repeat and compare:

flowchart LR
    IN([input]) --> T1["LLM: try 1"] & T2["LLM: try 2"] & T3["LLM: try 3"] --> V["compare / vote"] --> OUT([output])
  • Sectioning: split the input into independent chunks, process them concurrently, aggregate. Wall-clock time ≈ the slowest chunk instead of the sum.
  • Voting: ask the same question several times (or from several angles), then keep the majority or intersect the findings. Randomness becomes a feature: independent tries catch different mistakes.

A concrete example

Reviewing a pull request:

parallel:
  call 1 → "review ONLY for security issues"
  call 2 → "review ONLY for performance issues"
  call 3 → "review ONLY for missing tests"
merge   → deduplicate, sort by severity

Each focused reviewer catches more than one generalist reading everything — and the three run in the time of one.

When to use it

  • Sectioning: the chunks are truly independent (documents, files, categories) and volume or latency matters.
  • Voting: the cost of a wrong answer is high (hallucinations hurt), and you’ll happily pay 3× the tokens for confidence.

When to avoid it

  • Chunks depend on each other (part B needs part A’s result) → that’s prompt chaining.
  • Budget is tight: parallel calls multiply cost linearly.

The classic trap

The merge step. Splitting is easy; recombining well is the hard part — deduplicating findings, resolving contradictions between voters, keeping a coherent tone across sections. Budget as much care for the merge as for the calls themselves.