Patterns for getting multiple AI agents to work together without stepping on each other. Schema-first tools, typed contracts, and the coordination glue that keeps things from falling apart at scale. ## Install bash /plugin marketplace add 2389-research/claude-plugins /plugin install building-multiagent-systems

Architecture Each agent has four layers: reasoning (the LLM), orchestration

(coordination logic), a tool bus (permissions and access), and deterministic adapters (the testable execution boundary). The separation matters — you can swap or test each layer independently. ## Coordination patterns Seven ways agents can work together: fan-out/fan-in, sequential pipelines, recursive delegation (agents spawning sub-agents), work-stealing queues, map-reduce, peer collaboration, and the MAKER pattern for million-step zero-error execution. Which pattern to use depends on your problem. Fan-out for embarrassingly parallel work. Pipelines when each stage depends on the last. MAKER when you need sustained accuracy over thousands of steps. ## Running in production Permission inheritance so sub-agents don’t escalate. Cascading stops and orphan detection so nothing runs forever. Heartbeat monitoring across hierarchies. Checkpointing and cost tracking so you know what you’re spending.