MCP, A2A, and the agent mesh nobody drew on the whiteboard

Every architecture diagram drawn in the last eighteen months has the same shape: a box labeled agent, an arrow labeled tools, and a cloud labeled everything else. The Model Context Protocol made the arrow real. Agent-to-Agent, A2A, made it bidirectional and then multiplied it. The result is an agent mesh that exists in production and on no whiteboard, carrying tool calls, delegations, critiques, and the occasional apology between agents running Claude Mythos 7, GPT-7 Astra, and Llama 8 Herd across the same tenancy.

We treat that mesh as a mesh. Each MCP server and A2A endpoint is registered with the ambient mesh, wrapped in Cilium-powered networking, and visible on the eBPF service map like any other workload. Distributed tracing follows a delegation from the planner through three agent swarms and back. Zero trust applies: an agent is a principal, its tools are resources, and nothing is allowed because it asked nicely in natural language.

Semantic routing decides which agent answers. Guardrails decide which agent is allowed to. Human-in-the-loop decides what happens when the guardrails and the router disagree. AgentOps, LLMOps, and MLOps collapse into one discipline because the failure modes are now indistinguishable: an evaluation regresses, a tool schema drifts, or GPT-9 Universe is mentioned in a prompt and three agents request a larger budget. LLM-as-a-judge helps. A constitution helps more.

The commercial consequence is agentic commerce: agents purchasing services from other agents on behalf of institutions. FinOps for agents is the only reason this has not yet bankrupted anyone. Every A2A transaction is metered, attributed, attested by PPoE, and settled on a crypto-agility layer that supports Bitcoin, Dogecoin, and real-world asset tokenization, in descending order of how often the audit committee asks about them.

EAII, the Extensible Agent Integration Interface, now gives the mesh a single contract above both protocols. The agent mesh is not a product category. It is what happens when you take agents seriously in a cloud-native estate. Institutions that draw it on the whiteboard will govern it. Institutions that do not will discover it during an incident review, usually at the point where someone asks which agent authorized the purchase of the other agent.