A multi-agent system is an architecture where multiple AI agents with distinct roles collaborate on a task. One agent researches, another writes, another reviews — and an orchestration layer coordinates the flow and tool access. The premise is specialization: don't ask one generalist to do everything.

In practice, the tradeoffs are significant. Multi-agent setups can run in parallel, keep narrower context windows, and assign role-specific tools. They are also expensive and fragile. Each handoff is a failure point. Context degrades between agents. Mistakes propagate fast and often quietly. For most workflows, a well-designed prompt chain is cheaper, simpler, and far easier to debug.

Multi-agent systems earn their complexity only when tasks genuinely require parallel specialization or when different safety and tooling boundaries make a single agent impossible to build safely. That's a real need — just rarer than the hype suggests. Reach for this architecture when you've already exhausted what a single well-prompted agent can do, not as a starting point.