MCP — the Model Context Protocol — is the connective tissue between my agents. In my system, sub-agents don't talk directly to each other. They expose tools through MCP and the orchestrators route requests by intent across the MCP layer.
The shape of the problem is integration, not intelligence. An agent that can search the web, query a database, post to Slack, and check the calendar is doing four different I/O jobs. Without a protocol, each integration is bespoke — a custom function call, a different auth shape, a different error model, a different retry policy. With a protocol, every tool looks the same to every agent.
MCP doesn't make agents smarter. It makes them composable. The orchestrator can route a request to whichever sub-agent has the right MCP server attached, and the sub-agent can call any of its tools without knowing how they're implemented underneath. New tools enter the system by adding an MCP server; new agents enter the system by declaring which servers they need access to.
The architectural payoff is that the messy parts — credentials, rate limits, retries, telemetry — live in the MCP servers, not in the agent prompts. The agents stay narrow. The plumbing carries the complexity.
This is the boring, load-bearing infrastructure of the Hermes system. Most of what makes the agents useful is what they're allowed to do, and MCP is the layer that decides.