--- title: "The protocols settled. The processes did not." date: 2026-09-02 url: https://remiam.co.uk/notes/protocols-settled-processes-did-not tags: [AI, Engineering, Operations] read_time_minutes: 6 description: "A2A joined the Agentic AI Foundation alongside MCP on 17 August 2026, settling the agent interoperability question. Deloitte puts agents in production at 11 per cent. Why the gap is operational, not technical, and how we scope around it." --- # The protocols settled. The processes did not. *Published 2026-09-02 · 6 min read · by Liam (Remiam)* On 17 August the last big agent protocol moved into the same neutral foundation as the other one. The interoperability argument is basically over. Meanwhile only about one company in nine has an agent actually running in production, and that gap is not a technology problem. On 17 August 2026, Agent2Agent became a hosted project of the Agentic AI Foundation at the Linux Foundation, sitting alongside the Model Context Protocol, which had been one of the foundation's founding contributions when it launched the previous December. Each protocol keeps its own maintainers, its own specification process and its own release cadence (Axios, 2026). If you have spent the last eighteen months waiting for the standards fight to resolve before committing to anything, that is your signal. It has resolved. One neutral home, two protocols, no vendor holding the pen. So the plumbing is settled. Now look at what is actually running. ## Eleven per cent. Deloitte's Tech Trends 2026 puts the number of organisations with agents in production at 11 per cent. Thirty-eight per cent are piloting. Forty-two per cent are still writing a strategy, and 35 per cent have no strategy at all (Deloitte, 2026). Gartner expects 40 per cent of agentic projects to be scrapped by 2027, and its stated reason is not model quality: it is that companies are automating processes that were designed around people, rather than redesigning the process first (Gartner, via Deloitte reporting, 2026). Read those two paragraphs together and the shape of the year becomes obvious. The hard technical questions got answered by other people, for free, in public. The thing standing between a pilot and a production system is not a protocol. It is the same set of unglamorous questions that have blocked every operational rollout we have worked on since 2014. ## The questions a pilot never has to answer. A pilot runs in a room with the person who built it sitting next to it. Production does not. These are the questions we now put in the scope document before anyone writes an agent, because every one of them is a place a pilot quietly dies. - Who owns it on a Tuesday. Not the project sponsor, the named person who gets the alert when it does something odd at 14:00 and has the authority to switch it off. - What it is allowed to touch. Write the permission list first, in plain language, and make it narrower than feels comfortable. Widening later is a conversation; narrowing after an incident is a crisis. - What happens when it is wrong. Not if. The error path needs a human landing point, a way to see what the agent saw, and a way to undo. If the only recovery is a database restore, it is not ready. - How you know it is working. A number on a dashboard that someone checks weekly. Completion rate, escalation rate, time saved per case. Vibes are how a pilot gets funded and how a production system gets quietly abandoned. - What it costs per run at ten times the volume. Pilots are cheap because they are small. Model the unit economics at real load before you commit, not after. - Whether the process should exist at all. This is the Gartner point and it is the one people skip. Half the workflows we are asked to automate were shaped around a form, a rota or an approval that stopped making sense years ago. ## Redesign beats automate. The last one deserves more than a bullet. When a client asks us to put an agent on a process, the first pass is always to sit with the process and count the steps that exist only because a human was doing it. Steps that exist to hand work between two people. Steps that exist so somebody can check somebody else. Steps that exist because the system could not do the thing in one go, so it was split. Automate that as-is and you get a faster version of a bad shape, plus a new thing to maintain. Cut it down first and quite often what is left is small enough that the agent is a thin piece of it rather than the whole project. That is a better system and a cheaper one, and it is also the version that survives its first bad week. > The protocol question was hard and it is now answered. The operational question was always the expensive one and it is still yours. ## What we would do this quarter. Stop treating standards risk as a reason to wait, because it is no longer a real risk. Pick one process with a clear owner and a countable output. Map it, cut the steps that exist for human handover, and only then decide what an agent does with what is left. Ship it narrow, instrument it from day one, and give it a name and a person. One working system in production teaches an organisation more than nine pilots. The numbers this month suggest most companies are still collecting pilots. ## References 1. [Exclusive: AI agents inch toward interoperability, Axios (2026)](https://www.axios.com/2026/08/17/a2a-agentic-ai-foundation-open-ai-standards) 2. [Linux Foundation Launches the Agent2Agent Protocol Project, Linux Foundation (2025)](https://www.linuxfoundation.org/press/linux-foundation-launches-the-agent2agent-protocol-project-to-enable-secure-intelligent-communication-between-ai-agents) 3. [The agentic reality check: Preparing for a silicon-based workforce, Tech Trends 2026, Deloitte (2026)](https://www.deloitte.com/us/en/insights/topics/technology-management/tech-trends/2026/agentic-ai-strategy.html) 4. [Deloitte Tech Trends 2026: Moving from AI Experimentation to Enterprise Impact, DigitalCxO (2026)](https://digitalcxo.com/article/deloitte-tech-trends-2026-moving-from-ai-experimentation-to-enterprise-impact/)