Parallelization: Sectioning & Voting
Splitting work across simultaneous calls, or asking many times and voting.
Parallelization runs multiple LLM calls simultaneously — in two distinct flavors this split-screen animation contrasts side by side.
Sectioning (left): one document fans out to three worker nodes, each doing a different job on it — answering, guardrail-checking, fact-extracting. Progress rings run at the same time, and the results zipper back together in an aggregator. Use it when subtasks are independent — no worker needs another's output — and each call performs better with focused attention on one concern than one call juggling all three.
Voting (right): the same question fires at five identical nodes. Their answers drop into a tally, and the majority wins. The nodes match on purpose — you're not dividing labor, you're buying independent perspectives to reduce the variance of any single call. Adjusting the acceptance threshold tunes false positives against false negatives, which is why voting suits reviews and safety checks where confidence matters more than cost.
Watch for the key visual difference: differently-labeled workers on the left, identical workers on the right. Same fan-out shape, opposite purposes — and remember the fan-out is fixed in advance. What happens when it can't be is the next pattern.