Multi-agent systems
Orchestrator-worker, role-based crews, swarm handoffs, debate, and Google's A2A protocol — when one agent isn't enough.
“One agent doing everything” hits a ceiling when the task structure is heterogeneous — researchers vs writers, planners vs implementers, generalists vs specialists. Multi-agent systems decompose the task across cooperating agents.
The trap, repeated everywhere: multi-agent is more complexity, more failure modes, more cost. Earn it.
This module is being expanded.
Coming in the next revision:
- Orchestrator-worker. One agent decomposes, dispatches, aggregates. The dominant production pattern.
- Role-based crew. Agents with personas (researcher, analyst, writer, reviewer). CrewAI’s model.
- Swarm / handoffs. OpenAI’s pattern, now in the Agents SDK — lightweight agents pass control based on who’s best-suited.
- Debate / consensus. Agents argue or vote. For high-stakes reasoning.
- A2A (Agent-to-Agent) protocol. Google’s 2025 protocol — the “MCP for agent-to-agent communication.” Framework-agnostic discovery and messaging.
- The trap of premature multi-agent. “This would be easier with multiple agents” is often code for “the single agent doesn’t have the right tools yet.”