Vol. III · Single-Agent Patterns · III.3
Parallelization
Fan-out independent LLM work, then reduce. Sectioning for speed; voting for diversity. Aggregation is code, not another creative writer.
5 min read
Doctrine
- Parallelize only independent subtasks. Shared mutable context is a race.
- Voting helps when the metric is 'did two of three reviewers agree'.
- Guardrail models run in parallel with the main call — do not block generation on a serial safety hop if you can overlap.
Sectioning versus voting
Sectioning: split a report into chapters, write in parallel, stitch. Voting: three reviewers score a patch, majority wins. LangGraph's Send API and CrewAI async tasks implement the former; eval harnesses implement the latter.
Anti-patterns
- Spawning parallel agents that all write the same file.
- Voting with temperature 0 three times and calling it ensemble.