When one agent's context, tools, or focus aren't enough, the instinct is to split the work across several specialized agents. Sometimes this is exactly right; often it is a costly mistake that multiplies tokens, latency, and failure modes while adding little. This chapter covers the topologies that work, the protocol that lets agents from different vendors interoperate, and — just as important — when not to go multi-agent.

12.1 Topologies of collaboration

Multi-agent designs differ in how control and information flow between agents. The dominant pattern is orchestrator–worker: a lead agent decomposes a task and delegates subtasks to specialized workers, then synthesizes their results — a direct application of the fan-out/fan-in primitive from Chapter 4. Hierarchical designs nest this across several layers for complex domains. Blackboard (shared-state) designs let agents collaborate by reading and writing a common workspace rather than messaging directly. Sequential pipelines pass work agent-to-agent in stages.

12.2 The discipline of not going multi-agent

Each additional agent consumes its own context and token budget, adds a coordination and failure surface, and makes the system harder to evaluate and debug. Published experience (including Anthropic's own multi-agent research) is consistent: multi-agent systems can dramatically outperform single agents on complex, parallelizable, breadth-first tasks — and they burn many times the tokens to do it. They are a poor fit for tightly coupled tasks where subagents need each other's intermediate results, or for anything a well-engineered single agent with good tools and context already handles.

12.3 Agent-to-Agent interoperability (A2A)

MCP (Ch. 8) connects an agent to tools; it does not standardize how two agents — possibly built by different teams or vendors, each opaque to the other — collaborate. The A2A protocol (originated at Google, now under the Linux Foundation) fills that gap. Each agent publishes an Agent Card advertising its identity, skills, and endpoint; other agents discover it and exchange tasks and results over a defined HTTP-based protocol, supporting long-running tasks and streaming. The mental model: MCP is how an agent uses a tool; A2A is how an agent delegates to a peer. A mature platform speaks both — MCP downward to capabilities, A2A outward to other agents.