Every agentic system we've built over the last 18 months ran into the same wall: integrating models with internal DoW systems is a one-off engineering project every time. Service A wants OAuth; service B wants mutual TLS with a hardware token; service C is a SOAP endpoint that hasn't been touched since 2014. The model is the easy part. The plumbing is the contract.
Anthropic's Model Context Protocol, MCP, changes the shape of that problem. Instead of writing N adapters from N agents to M services, you stand up M servers that speak MCP, and any agent that speaks MCP can use them. Same idea as LSP for editors, but for tools and context instead of language servers.
Why this matters for federal: most DoW systems will never be rebuilt to be AI-native. They will be wrapped. MCP gives you a single, auditable wrapper format. The wrappers themselves can sit inside the same accreditation boundary as the services they front, so the agent never crosses an enclave boundary to reach a tool.
The other underappreciated property: MCP servers are inspectable. A reviewing engineer can list the tools an agent has access to, see the exact schema of each, and review the audit log of which tools were called with what arguments. That maps cleanly onto RMF artifact requirements and acceptance review.
What we're doing on our programs right now: every internal Nexagen tool that an agent might want to call gets an MCP wrapper before it ships. Function-calling against bespoke schemas is now considered tech debt on day one.
The catch is that MCP is moving fast. The spec at the moment is roughly six months from a 1.0 surface that we'd feel comfortable promising on a contract deliverable. For pilots, it's the right answer. For long-lived production, our recommendation is to keep the orchestrator's tool layer behind a thin adapter so you can swap MCP versions without re-engineering the agents.