MCP is not exciting because it makes agents suddenly intelligent. It is important because it makes agent infrastructure less accidental. Tools, resources, and prompts become contracts rather than one-off adapters hidden inside application code.
The maintainability advantage is substantial. If the model changes, the retriever changes, or a tool backend moves, the agent should not require a full rewrite. Typed interfaces also make testing possible: the team can validate arguments, permissions, outputs, and failure behavior separately from the model.
The long-term value of MCP is institutional, not theatrical. It reduces the cost of changing components while increasing the visibility of what the agent is allowed to do.