Most agent architecture arguments dissolve once you ask a narrower question: who decides what happens next? These five patterns are the useful answers, from most deterministic to least.
01Prompt chainingSplit one hard task into ordered steps, each with its own prompt and its own check. Use when the task decomposes cleanly; avoid when steps are independent, since chaining only adds latency.
02RoutingClassify first, then send to a specialist with a narrow prompt and a small tool set. Use when inputs fall into distinct kinds; avoid building the router on an LLM when a classifier is cheaper and more accurate.
03ParallelizationFan out independent work and merge. The only pattern that makes a pipeline faster rather than slower. Avoid when later steps need earlier output.
04Agent loopThink, act, observe, repeat — with a hard iteration cap and a timeout that are not optional. Use when the step count is genuinely unknown.
05Structured outputMake the model fill a schema rather than write prose you then parse. Validation failure is a retry, not a crash. Use anywhere output feeds another system.
Choosing between them
Work down the list and stop at the first pattern that can do the job. If you can enumerate the steps on a whiteboard, you do not need the agent loop. If the steps have no dependency on each other, parallelization beats chaining. If the output is consumed by code, structured output is not optional regardless of which other pattern you picked.
