Previously at the Volcano Base I turned part of myself into a download. Since then, school half-term has made project progress slow but, thankfully, steady.

Somewhere beneath your feet right now, in a metaphorical sense that doesn't involve actual excavation, the plumbing of the agentic web is being installed. Not in a press conference or a TED talk. Quietly, in GitHub repositories and W3C community group meetings and Linux Foundation working documents, with the unhurried confidence of people who know they are building something foundational and would rather you didn't ask too many questions while they're doing it.

The web exploded in the 1990s for a lot of reasons, but one of them was boring and important: standards. Tim Berners-Lee founded the W3C in 1994, having already noticed that competing implementations were beginning to fragment the web he'd just invented. The idea was to establish neutral, consensus-driven governance before things got truly chaotic. It was, in retrospect, an unusually sensible thing to do.

The browser wars that followed demonstrated what happened when that governance didn't extend far enough. The lesson was simple: whoever controls the standards controls the infrastructure. Whoever controls the infrastructure shapes everything built on top of it.

Three things happened in the last few months that are all connected.

First, the Model Context Protocol. MCP is the standard that lets AI agents connect to tools and data sources reliably, without everyone having to build their own bespoke connectors. Think of it as the universal cable that means your AI assistant can actually talk to your calendar, your CRM, your database: without a small army of developers writing custom code for each one. Anthropic created it. Then, in December 2025, they donated it to the newly formed Agentic AI Foundation, a directed fund under the Linux Foundation, co-founded with OpenAI and Block, and backed by Google, Microsoft, AWS, and Cloudflare. The protocol is now, formally, nobody's property.

Second, Agent Skills. Last week I turned myself into a download. What I didn't mention was that the format I used (the Skills specification) was simultaneously being opened as a public standard. Anthropic published the spec. OpenAI adopted it for Codex. Skills.sh launched as a community directory with over 65,000 skills and counting. A Skill is, in essence, a package containing a piece of professional expertise that any AI agent can pick up and use. It's becoming the standard way to package procedural knowledge for agents, and it's already cross-platform.

Third, WebMCP. On 10 February 2026, the Chrome team quietly published a blog post about something called WebMCP. It's a proposed web standard, co-authored by Google and Microsoft engineers, being incubated through the W3C's Web Machine Learning Community Group. The idea is straightforward: instead of AI agents squinting at your website like a confused tourist trying to read a menu in a foreign language, your website explicitly tells agents what it can do and how to do it. A structured tool layer for the agentic web, right in the browser. If an AI assistant wants to book a flight on your behalf, rather than fumbling around clicking things and screenshotting results, WebMCP lets the airline's website say directly: here are the tools I offer, here's how to use them.

Here is the version for everyone who skimmed the last three paragraphs: the internet is getting another layer.

We’re familiar with the layer you can see: websites, apps, interfaces. It was built for humans with eyes and time and the ability to interpret a navigation menu. The new layer is being built for agents: structured, machine-readable, explicitly defined.

This matters to you for a practical reason. The decisions being made right now about what these standards look like, who governs them, and how open they remain will determine what AI agents can and can't do with the digital surfaces your organisation uses every day. Your website, your internal tools, your supplier portals. If those surfaces speak the new language, agents will be able to interact with them reliably and efficiently. If they don't, agents will be fumbling around guessing, which is slower, more expensive, and prone to error.

The organisations that understand this early will be able to think sensibly about their own digital estate. Not with urgency, not with panic, but with the quiet advantage of having noticed something before it became obvious.

It would be slightly dishonest to present all of this as an unambiguous triumph of open governance. Developer Gregory Golberg made a pointed and entertaining argument that WebMCP running through the browser means the companies behind the dominant browsers control a rather important chokepoint.

Chrome is the dominant browser. Microsoft Edge (Chromium-based since 2020) is also built on the same open-source engine that powers Chrome. Which means the two most widely used browsers on the planet share the same underlying codebase, and Google employs the vast majority of the people who develop it. Chromium is technically open source, but "open source with one company steering" is a familiar story in tech, and it doesn't always end with the steering wheel being handed over voluntarily.

Golberg compared WebMCP to Google's AMP initiative: technically open, developed under a standards banner, but structured in a way that required routing through Google's infrastructure. The W3C process is reassuring, and Microsoft co-authoring the spec helps. But Firefox and Safari haven't committed to anything yet, and a standard that only two Chromium-based browsers implement is not quite the same thing as a standard.

The Linux Foundation has genuine form in neutral stewardship. The W3C process, when it works, is genuinely consensus-driven. But it's worth noticing that the companies setting these standards are the same companies with the most to gain from how they're set. This has always been true of technology infrastructure, and it has never stopped being worth mentioning.

This isn't a story about a problem waiting for a solution. The governance question (the one that should keep enterprise lawyers up at night) was quietly answered three months ago.

The Agentic AI Foundation already exists. It was announced in December 2025. Most people missed it. The plumbers just didn't issue a press release that anyone read.

Have a poke around skills.sh and ask yourself: is there anything your organisation knows professionally that could live in a zip file?