Mastering AI Agent Architectures: Flow Styles

Stop an error where it starts, or watch it travel.

Flow styles move work through stages: chains, topics, logs, streams and sagas. This book measures fifteen of them on the Gauntlet benchmark and shows how the flow family controls propagation — the way one stage’s mistake becomes the next stage’s input.

Pipes and filters — a filter between stages lifts research success from 63 % to 92 %Backpressure and rate control for agent fleetsPub-sub teams and event sourcing as team memoryStreaming cognition and its latency budgetSagas and compensation — the lowest plausible wrongness on transactions in the seriesDelivery semantics — at-most, at-least and exactly-once, measured under faultsChoreography versus orchestrationEach style is read on four axes and under node failure; the decision card states when flow wins and what it costs on the clock.

Runnable code, frozen results and a lock at tag gauntlet-b4. Book 4 of 7 in Mastering AI Agent Architectures — The Styles Series.