The Protocol Layer Settled
The agent interoperability wars are basically over, and they ended the way these things usually do: with a standard nobody fully owns. The Model Context Protocol crossed 97 million monthly SDK downloads, runs on more than 10,000 enterprise servers, and has native support from Anthropic, OpenAI, Google, Microsoft, and AWS. In December 2025, Anthropic donated it to a directed fund under the Linux Foundation. That's the move that settles it. A protocol governed by a community is a protocol you can build a company on.
Two protocols, two jobs
The pattern most enterprise architectures are converging on in 2026 uses two layers, not one. MCP handles how an agent reaches tools and data. A2A handles how agents delegate work to each other. Individual agents pull context through MCP. Task handoff between agents goes over A2A.
It's a clean split, and it maps neatly onto how finance work already flows. An orchestrator breaks a close into pieces, hands each piece to a specialized agent, and each of those agents reaches into the systems it needs through a governed connector.
Why standards are good news for the unglamorous middle
When the plumbing is contested, every integration is a custom build and a renegotiation. When the plumbing is standard, integration becomes a connector you write once. That is exactly what enterprise finance needs, because the value was never in the protocol. It's in the systems the protocol reaches: D365, SAP, Workday, the data lake, the places where the actual money lives.
A settled standard does not make MCP the moat. It makes the moat the thing it always was, the governed connection between an agent and a system of record that someone spent fifteen years configuring.
One caution worth keeping in view. Gartner still expects more than 40% of agentic AI projects to be cancelled by 2027, mostly over unclear value and weak governance. A standard protocol removes one excuse. It does not remove the hard part. It just means that when you fail now, you can't blame the plumbing.