Chat is the new UI? What we took away from AGNTCon + MCPCon Europe 2026

20.09.2026 - David Bongard in Development

In mid-September, developers, platform builders and companies that no longer just experiment with AI agents but run them day to day gathered in Amsterdam. AGNTCon + MCPCon Europe took place on 17 and 18 September at RAI Amsterdam, organised by the Agentic AI Foundation under the umbrella of the Linux Foundation. We were there to understand where the field is heading – away from the hype headlines and product announcements.

A caveat up front: not everything we heard will matter tomorrow. But several developments touch quite directly on what we work with every day: websites, web applications, and the question of how people (and, increasingly, machines) use them.

In short: MCP, and why there’s a foundation behind it

The Model Context Protocol (MCP) is an open standard that lets AI applications such as Claude or ChatGPT talk to external tools, data and services. Think of it as a shared language an AI assistant uses to ask an application “What can you do?” and then “Please do that.”

MCP was originally developed by Anthropic. In December 2025, Anthropic handed the protocol over to the Agentic AI Foundation (AAIF), a fund under the Linux Foundation co-founded with Block and OpenAI and supported by Google, Microsoft, AWS, Cloudflare and Bloomberg. Alongside MCP, the foundation also hosts the agent framework goose and AGENTS.md, a standard that gives coding agents project-specific guidance for a repository (Linux Foundation).

Why does this matter? The opening keynote drew a comparison with the power grid: if agents are going to handle transactions and decisions in future, the infrastructure behind them shouldn’t belong to a single company. For agencies like ours, neutral governance mainly means planning security. Building expertise in MCP is a bet on a standard rather than on a vendor.

A new door: websites that offer tools to agents

The topic that excited us most was WebMCP. Today, AI agents “operate” websites by reading pages, interpreting screenshots and hunting for buttons. That’s error-prone and expensive. WebMCP turns this around: a website publishes structured tool descriptions, and the agent calls them with clearly defined parameters instead of feeling its way through the page.

The declarative variant is particularly elegant: an existing HTML form simply becomes a tool. An example from one of the talks:

<form toolname="createSupportTicket"
      tooldescription="Create a support ticket"
      toolautosubmit>
  <input name="issue">
</form>

Name, purpose, parameters: that’s all an agent needs to understand what this form does. The attributes toolname, tooldescription and toolautosubmit register actions that AI agents can call in the browser (Web Features Explorer). toolautosubmit determines whether the agent may submit the form itself or whether a human confirms. Unlike classic MCP, WebMCP runs entirely in the browser: the tools use the session the person is already logged into, and the browser governs what the agent is allowed to do.

Josho Brouwers and David Bongard at AGNTCon + MCPCon Europe 2026

An honest assessment belongs here too: WebMCP is still very early. The specification is published by the Web Machine Learning Community Group and is neither a W3C standard nor on the standards track. Chrome currently has the only working implementation, behind a feature flag; Microsoft co-authored the spec, while Mozilla and Apple have published no timelines. There is also criticism of the process: members of the community group report learning about WebMCP from press coverage and ask whether the Chrome implementation will become the de facto standard regardless of community input (WebMCP Technical Notes).

Still, the idea that a website offers a clean interface for agents alongside its interface for people is likely here to stay, however the details end up looking.

The last mile: who designs what users end up seeing?

The keynote “Three Doors to One Tool: MCP vs WebMCP vs CLI” by Frédéric Barthelet (Alpic) and Dominic Farolino (Google) focused on the last mile – the moment a result reaches the person. They distinguished three variants:

No interface: the agent acts on its own, and the person gets a text message at best.
Aggregated UI: the chat platform assembles its own presentation from various sources, such as a recipe card.
Branded UI: the provider delivers its own interface directly into the chat, with its own look and its own logic.

For brands, the difference is considerable. Without an interface of your own, you become an interchangeable data source for the platform.

The third variant is made possible by MCP Apps. Since January 2026, MCP Apps has been the first official extension of the protocol: tools can display interactive interfaces such as dashboards, forms or multi-step workflows directly in the conversation. These interfaces run in a sandboxed iframe, and every action triggered from the interface goes through the same review and consent path as a direct tool call (MCP Blog). It is now supported by ChatGPT, Claude and GitHub Copilot Chat in VS Code, among others. How different approaches to such agent interfaces (AG-UI, A2UI, MCP Apps) fit together was the subject of a dedicated talk by Manfred Steyer of Angular Architects.

We took one formula home from the keynote day: UX = AX + UAX. Agent Experience (AX) describes how well an agent understands and can operate a product – clean interfaces and tools. User-Agent Experience (UAX) describes what the human ultimately sees, understands and decides. One slide put it neatly: “Headless, but with a human eye at the end.” In other words: even if the website is no longer the only way in, a human always looks at the result in the end and has to be able to trust it. That is a design task as much as a development one.

And WordPress?

As a WordPress agency, we naturally paid attention to where our most important system stands. The good news: WordPress is moving in the same direction. Since version 6.9, the Abilities API has been part of WordPress core – a kind of registry in which plugins and themes describe what they can do. The MCP Adapter, maintained by the WordPress AI team, translates these abilities into MCP tools that external AI agents can discover and call (Learn WordPress).

We like the cautious default: abilities are not reachable via MCP unless they are deliberately marked as public. Which functions of a website are made available to agents is therefore an active decision, not a side effect of a plugin update.

Beyond technology: how agents are changing development work

A second thread running through the conference concerned not the websites themselves but the work on them. The talk “Pull Requests Are Dead, Long Live Peer Review” by Dylan Ratcliffe (Overmind) summed it up: when coding agents take over implementation, human work shifts forward – to research, conception and planning. Review moves from the finished code to the plan. People decide on the approach and the trade-offs up front, and the plan becomes the yardstick against which the implementation is checked. Overmind is building a tool for this called Until, which compares an agent’s implementation against the agreed plan until every material difference has been resolved.

Other talks added the organisational perspective. Booking.com described how early MCP experiments grew into company-wide infrastructure. In “90 Days To Agentic Engineering”, Thomas Schöne talked about how teams capture rules for agents on a per-project basis. And to close, Liran Tal (Snyk) reminded the audience that code written by agents needs deterministic checks too, for example automated security scans built firmly into the workflow. More speed doesn’t replace quality assurance – if anything, it makes it more important.

Outlook: what we can take from this

So what are we taking away? A few cautious theses:

Websites are gaining a second audience. Alongside people, agents are becoming visitors. Structured data, clean forms and clear interfaces are likely to become as important as search engine optimisation did twenty years ago.
Standards are consolidating, but slowly. MCP is under neutral governance and maturing. The upcoming 2026-07-28 specification brings, among other things, a stateless protocol core, an extensions framework and a formal deprecation policy. WebMCP, on the other hand, is still an experiment, and whether Firefox and Safari will follow remains open.
Brands are moving into the chat. Those who only supply data hand the design of the experience over to the platform. Owning an interface inside the chat therefore becomes a design question.
The website isn’t going away. “Chat is the new UI” is a strong phrase, but it doesn’t apply equally to every audience. In B2B settings or among older users in particular, the classic website will remain the main channel for a long time.
Open protocols are no guarantee against concentration. The open web produced large platforms too. Neutral standards don’t prevent power from accumulating one level up, for example with the providers of the major agents.

We’ll keep following these developments, and we’ll certainly try some of them out. Session recordings are expected to become available after the conference; if you want to dig deeper, the full programme is on the conference’s Sched page.

If you’d like to know what these developments could mean for your website or web application, get in touch.

Sources and further reading

About the author - David Bongard

David Bongard has been building websites since 1998 – first as a web developer, and since 2014 as managing director and CTO of Resonanz Digital in Vienna. Together with his team, he develops custom web applications and e-commerce solutions for companies with demanding digital requirements – from concept and implementation through to ongoing support in live operation.

Back to overview