Agent patterns
More agents does not mean smarter. Every orchestration shape buys a capability and introduces a new way to be wrong. This separates the common shapes: what each looks like, when to use it, when not to, and how it fails.
- 11
- patterns
- 3
- with a citable origin
- 33
- failure modes documented
Sequential (pipeline)
no single origin on recordA's output is B's input, all the way down.
Linear, one-way, no回頭路
3 failure modesParallel (fan-out / fan-in)
2025-06-13The same input goes to several workers at once; results are merged at the end.
One to many, then many to one
3 failure modesIterative (loop)
2022-10-06Act → observe → act again, until it converges or the budget runs out.
A cycle with a feedback edge
3 failure modesRouter / dispatch
no single origin on recordClassify the request first, then hand it to the matching handler.
One in, many out, only one path taken
3 failure modesOrchestrator–worker
2025-06-13A lead decomposes the task, dispatches workers, and collects results.
Hub and spoke; the lead holds the whole picture
3 failure modesEvaluator–optimizer
no single origin on recordOne produces, one scores; below threshold goes back for another pass.
A closed loop between producer and evaluator
3 failure modesReflection
no single origin on recordAfter producing, criticise the output once, then revise.
Two passes of one agent
3 failure modesDebate / adversarial
no single origin on recordSeveral positions attack each other; disagreement surfaces the weak points.
Opposing agents, a third adjudicates
3 failure modesShared blackboard
no single origin on recordSeveral agents read and write one shared state instead of messaging each other.
Shared store plus several independent readers and writers
3 failure modesEvent-driven
no single origin on recordWoken by external events rather than by a person asking.
Watch → trigger → act → report
3 failure modesHuman-gated
no single origin on recordPause for human approval at the steps that matter; automate the rest.
Decision points inserted into an automated flow
3 failure modes
Not sure which to start with?
Most tasks need only a pipeline or a loop. Before adding an orchestration layer, ask: can this be done in one pass? If not, is the feedback signal reliable? Only when both answers are clear is orchestrator–worker or fan-out worth it.