The Three Ways People Build AI Agent Systems: From Fixed Workflows to Concurrent Collectives

A taxonomy of three architectural patterns for AI agent systems: sequential chains (Agentic Waterfall), parallel DAG graphs with fan-out/fan-in, and event-driven concurrent collectives. The post argues that chains serialize unnecessarily, graphs still block at join barriers and require pre-planned structure, while event-driven runtimes let agents emit and subscribe to semantic events — enabling true concurrency and mid-flight adaptation. The author illustrates the thesis with a real incident where their planner agent was killed by a wall-clock watchdog because it lived outside the event-driven runtime, blind to execution feedback. Moving the planner into the shared event bus gave it real admission receipts, live execution news, and output-based supervision. The post introduces Mozaik, an open-source TypeScript runtime built around this model, and argues that genuine intelligence in agent systems requires real concurrency — not just parallelism between predefined steps.