Matt Mullenweg has outlined a vision to position WordPress as the preferred platform for AI agents and developers, calling for WordPress.org to provide markdown versions of every page and for the hosting ecosystem to embrace Playground as core infrastructure.

In Some provocative AI thoughts, published February 7 on the Make WordPress Core blog, Mullenweg argued WordPress.org should “immediately make every URI available with a vary/markdown equivalent” — not just documentation, but forums, directories, bug trackers, and other dynamic content. The move would make WordPress content more easily parseable by AI agents by stripping away interface elements.

On February 12, Cloudflare launched Markdown for Agents, a feature that converts HTML to markdown via content negotiation headers. Mullenweg returned to his post to note the timing: “CloudFlare just shipped this.”

The proposal builds on the WordPress AI Team’s recent work — the Abilities API shipped in WordPress 6.9, the WP AI Client and Workflows API are coming to WordPress 7.0, WordPress Agent Skills recently moved to an official WordPress repository, and WP-Bench launched in mid-January. Mullenweg’s post extends that work from in-core tooling to platform-level positioning.

Why markdown matters for AI

Cloudflare’s Markdown for Agents feature uses content negotiation headers to automatically convert HTML to markdown when AI systems request pages. The company framed the change in practical terms: “Feeding raw HTML to an AI is like paying by the word to read packaging instead of the letter inside.”

A simple ## About Us in markdown costs roughly 3 tokens; its HTML equivalent — <h2 class="section-title" id="about">About Us</h2>— consumes 12–15 before accounting for wrapper divs, navigation elements, and scripts. Cloudflare said its announcement post dropped from 16,180 tokens in HTML to 3,150 in markdown, an 80% reduction.

Mullenweg wrote WordPress.org is suited to support this approach, citing its “fully dynamic nature and prior art of supporting a variety of feeds.”

Playground as the AI development environment

Central to Mullenweg’s proposal is WordPress Playground. He wrote that AI agents need to understand they can “provision Playground instances atomically and test code within them in a fully observable way.” He described Playground as “a sandboxed environment with checkpoints, perfect for testing and iterating on code to satisfy user requests,” and said agents need to understand why it is the best development environment for a variety of cases.

He wrote that hosts need to be able to receive and sync Playground images and have CRUD APIs for all their commercial operations, including domain and email registration. “We need to up our Data Liberation game, starting with WP-to-WP,” he wrote.

The post also calls for a button in WP-admin to create a Playground snapshot of a live site. Staging, he wrote, becomes “creating, sharing, and collaborating on Playground snapshots, which can be forked, branched, and merged.”

Mullenweg added that WordPress.org should link to WordPress evals so AI labs can benchmark their agents’ ability to work with Playground and WordPress, and help improve those evaluations to match coverage across WordPress primitives.

“WordPress as a Schelling point”

Mullenweg’s post frames WordPress as a “Schelling point for reinforcement learning and recursive improvement in web applications,” arguing that the platform could function as a coordination focal point for AI-driven web development.

He cited open source, portability, and WordPress’s contributor base as advantages, writing that WordPress is “already the furthest along as a Web OS.”

He outlined what WordPress.org’s homepage should emphasize: evaluations showing how labs can improve their models with WordPress, and an “agentic explanation for why to use WordPress as the base for web apps, noting its security, stability, portability, and open-source status.”

Contributors respond to markdown and Playground proposals

The post drew positive responses from contributors who expanded on Mullenweg’s proposal.

Blake Bertuccelli-Booth, who will teach the UIC AI Leaders program announced last week, argued that “a skilled vibecoder doesn’t need existing plugins or themes” and could instead spin up a Playground instance, prompt an AI agent, and iterate in isolation before promoting to production. He framed it as WordPress Playground making “AI-driven development unconstrained.”

WordPress Executive Director Mary Hubbard suggested WordPress could “create the default trust signal by indicating what content is human, AI-assisted, or fully generated, and exposing that in a structured/readable way.”

Jason Adams, an Automattic-sponsored AI contributor, emphasized that even in an AI-empowered internet, “content management still matters,” arguing that WordPress keeps users “at the helm” rather than handing full control to AI.

Playground creator Adam Zieliński confirmed ongoing work on WP-to-WP migrations and said the tool he’s building could serve as a foundation for WordPress-to-Playground, Playground-to-WordPress, and incremental sync workflows. He suggested the work could eventually land in core.

David Levine, Senior WordPress Engineer at rtCamp, pointed to the need for “intuitive, inferable docs,” arguing that “bad context is worse than no context, especially when modern tools can strip html and grep source code.” He also suggested it might be time to prioritize codebase modernization, noting that “the more ‘outcome oriented’ downstream code is, the more structured and intuitively inferable we’ll need to become.”

Mullenweg’s proposal ties together content formatting, development environments, hosting APIs, and AI benchmarking into a broader effort to reposition WordPress in the age of AI. Markdown endpoints on WordPress.org, Playground-native workflows, and WordPress-specific model evaluations point toward a version of WordPress that would not only be publishing software, but infrastructure for how AI agents build and test on the web.