Coordinating Multiple AI Agents: Where Multi-Agent Systems Actually Break

A developer who orchestrates several AI agents running in parallel, including researchers, implementers, and reviewers, argues that failures usually stem from the coordination layer rather than the models themselves. One incident saw a sub-agent unexpectedly overwrite a shared progress board because the update API replaced entire elements by ID, silently erasing other workstreams. Another involved reviewer agents flagging legitimate work as permission violations because each reviewer only knew its own card's allowed operations. The author's fixes include restricting sub-agents from writing shared state, using namespaced IDs, re-reading after merges, and centralizing adjudication with full visibility of parallel scopes. The core takeaway is that orchestration is a system design problem rather than a prompt problem.