Summary: Cloudflare launched Markdown for Agents in beta on February 12, 2026, converting origin HTML to markdown at the edge when an agent sends Accept: text/markdown. The article reports an 80% token reduction for the announcement page and notes Content Signals for AI-use preferences. SEO critics raise cloaking concerns because the header reaches the origin and could trigger different content.
Markdown for Agents arrived from Cloudflare on February 12, 2026, as a beta. The mechanism fits in one sentence: an AI agent attaches an Accept: text/markdown header to its request, and Cloudflare pulls the origin HTML, converts it to markdown at the edge in real time, and sends it back. Nothing is published at a second address — this is HTTP content negotiation doing exactly what it was designed for, with one URL serving different representations depending on the request headers.
Scale is why anyone should care. With roughly 20% of the web standing behind Cloudflare, “make your site legible to agents” shrinks from a side project into a single dashboard toggle. As traffic drifts away from search engines toward AI crawlers and agentic browsing, Cloudflare is acting as if agents are first-class citizens — clients to serve, not nuisances to block.
Negotiating Content at the Edge
The plumbing is tidy. A response arrives with content-type: text/markdown plus a vary: accept header, which keeps the two variants from mixing in cache, and an x-markdown-tokens header carrying an estimated token count that an agent can immediately put to work sizing context windows or planning chunking. Content Signals rides along too (ai-train=yes, search=yes, ai-input=yes), declaring which AI uses the content is licensed for, and Cloudflare says custom Content Signals policies are planned for later.
Cloudflare switched the feature on for its own Developer Docs and blog first, and any enabled zone can do the same: activation sits in the dashboard under the zone’s Quick Actions, free for Pro, Business, Enterprise, and SSL for SaaS customers. Tools such as Claude Code and OpenCode already send the header, so many agent stacks get the benefit on day one without touching a line of code.
Where the 80% Token Savings Come From
Cloudflare’s arithmetic is persuasive. The announcement post itself weighs 16,180 tokens when served as HTML and 3,150 as markdown — a reduction of 80%. A basic “About Us” heading runs about 3 tokens in markdown; in HTML the tags alone consume 12 to 15 before a single wrapper div, nav bar, or script tag has been counted.
For anyone building agents, that reorders the economics. The scenario of pulling ten pages into a context window now comes with a token bill 80% lighter, and what lands is structured headings and paragraphs instead of prose polluted by layout noise.
The Cloaking Question
The day after launch, Search Engine Land ran its coverage under the headline “SEOs on alert.” Consultant David McSweeney named the risk that actually matters: the markdown Accept header gets forwarded to the origin, which means origin servers can detect agent requests and hand back content no human will ever see — hidden instructions, altered product data — in effect a shadow web reserved for bots. Jono Alderson’s critique went after the premise itself: “When you flatten a page into markdown, you don’t just remove clutter. You remove judgment, and you remove context.”
The search engines spoke up as well. Google’s John Mueller asked the obvious question: “Why would they want to see a page that no user sees?” Microsoft’s Fabrice Canel warned about double crawling: “Really want to double crawl load? We’ll crawl anyway to check similarity. Less is more in SEO!” Taken together, the engines most likely to ingest this markdown are also the ones reserving the right to check it against the page humans see — which is precisely where the cloaking question bites.
Cloudflare never answered the cloaking criticism head-on; its stated rationale stayed with efficiency and the shift toward AI-mediated content consumption. The part of the design aimed at defusing trust concerns is Content Signals, which puts the site’s licensing position into machine-readable form instead of leaving crawlers to guess.
Practical Moves for Site Owners
Three effects are worth acting on. First, documentation and content sites whose audience now includes AI agents have the cheapest agent-friendliness play available — there is no second markdown corpus to maintain because the edge generates it, and Google search results are, in theory, untouched. Second, crawler authors who send the Accept header should understand that the header reaches the origin, the origin can identify you, and your behavior has to survive a “two versions disagree” audit. Third, Cloudflare Radar added a content_type dimension showing which content types AI bots receive, handing teams a quantitative handle on agent traffic.
When finer control is needed, Workers AI’s AI.toMarkdown() and Browser Rendering’s /markdown endpoint offer programmatic routes. This is also Cloudflare’s second agent-infrastructure move in the space of a few days — GLM-4.7-Flash just landed on Cloudflare Workers — and the platform strategy is unmistakable: wherever agent traffic goes, that is a product line.