Overview
Parallelization reduces elapsed time by running independent parts of a task simultaneously instead of forcing one agent to complete every step in sequence. It can also improve quality by keeping each worker's context short and assigning research, extraction, or analysis to separate agents before a stronger synthesizer combines the results.
The most reusable parallel pattern is fan-out and fan-in. One orchestrator expands a problem into several independent work packages, sends those packages to multiple agents, and then feeds their outputs into a dedicated synthesis step.