The protocol layer arrived

For two years, connecting an AI assistant to a business's tools meant choosing a platform and living inside its idea of what a tool was. That is changing at the layer below the product. In July 2026 the Model Context Protocol published its largest revision since launch, and in August the Agent2Agent protocol joined MCP under the same foundation. Neither event is exciting to a salon owner. Both change what a salon owner can walk away with.

What changed in MCP

The 2026-07-28 revision makes the protocol core stateless: a server no longer has to hold a session open for every client, which is what made hosting an MCP server awkward at scale. It adds header-based routing and multi round-trip requests, so a server can ask the client for more information during a single call instead of forcing a dozen separate exchanges. The official SDKs for TypeScript, Python, Go and C# shipped alongside it.

For a small deployment the practical effect is boring and good: an integration can run on ordinary serverless infrastructure, and a tool written for one assistant has a better chance of working with the next one.

A2A grew up, then moved house

Agent2Agent, the protocol for one agent to hand work to another across company boundaries, reached version 1.0 in March 2026. By April the Linux Foundation counted more than 150 organisations involved. In August, A2A joined the Agentic AI Foundation as a hosted project — the same home as MCP. Two protocols that solve different problems (tools for an agent; agents for each other) now sit under one open governance structure instead of two vendor launch posts.

Why a business should care

Standards are a procurement lever. If your booking agent speaks MCP to your calendar and A2A to your accountant's system, replacing the model or the vendor is a swap, not a rebuild. If everything is a proprietary connector, the switching cost is the product. Ask any supplier a plain question: which parts of this are standard, and what happens to my integrations if we stop working together?

There is a second, quieter benefit. Open protocols get audited by people who are not trying to sell you anything. A stateless core and a published authorisation model are easier to reason about than a black box with a support email.

What it does not solve

A protocol is not a permission system, an audit trail or a reliability guarantee. MCP tells an assistant how to call a tool; it does not decide whether this user, at this hour, may issue a refund. A2A lets agents talk; it does not make their answers true. The hard parts remain where they were: scoped credentials, human approval for anything irreversible, logs a person can read, and an eval suite that catches regressions.

Standards lower the cost of leaving. That is exactly why they are worth insisting on — and exactly why they are not a substitute for the boundaries we wrote about last month.