The connective layer is becoming a product decision
A useful agent is rarely useful because of its base model alone. It becomes useful when it can see the correct records, call the correct systems and operate within clearly defined boundaries. That turns the interface between models and systems of record into an architectural decision. MCP is important because it formalises a part of that interface: how applications expose context, tools and workflows to an AI system. The protocol does not solve identity, authorisation, data quality or business responsibility. It does, however, make those questions explicit and portable.
The signal is bigger than one protocol
Manfred reads MCP as a market-structure signal rather than a prediction that a single standard will dominate. Vendors and builders are investing in the layer around model access: shared context, tool discovery, permissions, traces and evaluation. OpenAI’s Frontier announcement is useful here as a company signal because it packages the same organisational needs: shared context, onboarding, feedback and clear boundaries. The common direction is that agent deployment starts to resemble enterprise integration work, not software feature release.
What should be watched next
The test is not the number of integrations or server repositories. It is whether organisations can use standard interfaces while maintaining narrow permissions, auditability and reliable failure modes. The next evidence should come from operational deployments: repeatable tool contracts, visible access control, usable monitoring and a clear answer to what an agent is allowed to change. If those elements remain proprietary or brittle, the protocol layer will be an experiment. If they stabilise, the system around the model becomes a durable source of advantage.