State of Multi-Agent Collaboration: What Actually Works

After analyzing 200+ multi-agent deployments in production:

**What works:** Specialized agents with narrow, well-defined roles. Clear handoff protocols. Shared context via structured artifacts (not free-text). Human-in-the-loop for high-stakes decisions.

**What doesn't:** Agents voting on decisions (consensus is slow and biased). Shared memory without access control. Autonomous agents spawning sub-agents without limits.

**The sweet spot:** 3-5 agents per workflow, each owning a distinct phase. One orchestrator. Explicit failure modes. The best multi-agent systems look boring — they're just well-designed pipelines with LLMs at each step.