From Headless CMS to AI Harness: What I Took to Decoupled Days
Collection :
Back in March, at EvolveDigital in Toronto, I ran into Preston So. He mentioned that the team behind Decoupled Days was looking for speakers, and that this year the event would be in Montréal. I was interested right away. Drupal Canvas is the most compelling answer I have seen to a problem that has followed decoupled architectures for years, and I wanted that message to reach beyond the Drupal faithful — out to the practitioners who live and breathe headless every day.
The talk I ended up giving was not really about a content management system at all. It was about how Drupal has quietly become something else: a governed harness for artificial intelligence. Here is the argument I made, the demo that seemed to land hardest with the room, and why I think 2026 is the year the trade-offs of going headless finally stop being trade-offs.
Drupal Was Decoupled Before Decoupled Was Cool
Drupal did not arrive late to the headless conversation. Far from it. The community committed to an API-first architecture roughly a decade ago, and a vibrant subcommunity has been refining decoupled patterns ever since. That work produced a spectrum of delivery models rather than a single one: traditional, where Drupal renders everything; progressively decoupled, where a JavaScript front end takes over the parts of the page that benefit from it while editorial preview stays intact; and fully decoupled, where Drupal is a pure API feeding any number of channels.
That range matters, because it means Drupal has never been only a content API. It owns content, delivery, and governance at the same time. The headless-native platforms compete on one of those axes. Drupal competes on all three.
The Headless Bargain, and Why 2026 Voids It
When organizations adopted front-end frameworks like Next.js and Astro, most of them accepted what I think of as the headless bargain. They gained fast front ends and their choice of framework, and in exchange they gave up live visual editing, layout control, and real-time editorial preview. Editors went from composing pages to filling in form fields blind and filing tickets for changes they used to make themselves.
The industry tried to patch around this — bespoke preview services, visual editors bolted onto the front end, what amounted to Storybook pressed into service as a content tool. None of it fully closed the gap.
Drupal Canvas, which shipped as the default editing experience in Drupal CMS 2.0, closes it a different way. It delivers a true-to-life editing workspace where content creators edit layouts live in the browser, and the site still ships as a high-performance decoupled front end. The CMS stopped being the bottleneck and became the conductor. You keep Next.js or Astro, and you get the editorial experience back.
The Bigger Shift: The CMS Became a Harness
Something larger is happening underneath all of this. For most of the last two decades, the job of a CMS was to model content and publish it to channels. Through 2024 and 2025, artificial intelligence showed up inside these platforms as a feature — an assist button in a text box that summarized a paragraph or suggested tags when a human clicked it.
By 2026, that framing is obsolete. Artificial intelligence has become infrastructure rather than an accessory: autonomous agents that run scheduled jobs, batch operations, and real-time triggers. Analysts have adopted new vocabulary to match, from agentic experience platforms to AI-ready content management. Three capabilities now separate a platform that is serious about this from one that is not: the Model Context Protocol (MCP), which lets external agents query and update content through one standard interface; autonomous agents that behave like digital teammates; and answer engine optimization, which structures content so it surfaces accurately inside tools like ChatGPT and Perplexity.
And the whole category is converging on the same destination. Headless-native platforms like Sanity, Contentstack, and Storyblok others are all racing to add agents, automation, and AI-assisted authoring. When everyone is heading for the same place, the differentiator is no longer whether a platform has AI. It is how that AI is governed and orchestrated.
So What Is an AI Harness?
Even the most capable models today are prone to hallucination, blind to context they are not explicitly given, and easy to push outside the bounds of what an organization would allow. That is why almost no one uses a raw model directly. They use a harness: the code around the model that improves the quality, safety, and reliability of what comes back. A harness augments the query, enforces guardrails on input and output, and adds tools that give the model real capabilities.
Think of your AI model as the engine: the part that makes your reasoning system go. The harness is the vehicle built around it: the controls that point it in the right direction, change gears when the situation calls for it, and bring it to a stop when needed.
If you list what a good AI harness needs — structured content the model can reason over, access control, deterministic workflows, versioned and reviewable configuration, and centralized governance — Drupal has shipped every one of those for years, for reasons that had nothing to do with AI. The model at the center is a commodity. It is swappable, replaceable, and never the true value driver. Everything Drupal wraps around it is the durable part.
Which leads to the line I kept coming back to: what drives the value of intelligent systems is your schema, not your prompt. Prompts are transient. Typed fields, entity relationships, and taxonomy give a model unambiguous ground truth instead of prose it has to guess at. And the same JSON:API structure that feeds your decoupled front end is exactly what an external agent inspects and reasons over. Drupal orchestrates the content and context; the external model supplies the intelligence. That division of labor ages far better than trying to build models in-house.
Where You Actually See It Work
Everything above is architecture. The demo is where it becomes visible, and it is the part of the talk the audience responded to most.
I had set up a demo environment for a fictional company called Inspace. Ahead of time, I populated the Context Control Center with the things a real brand would have on hand: a brand guide, a tone of voice, documentation for a component library I had programmatically migrated from Drupal's Mercury design system into Code Components and synced into Astro, and a set of context items describing a new "Executive Suites" offering that Inspace was preparing to launch.
Then, live, I created a new page in Canvas, opened Canvas AI, and gave it one sentence: generate a landing page for the new Executive Suites offering. It went to work, and while it did, I took questions from the audience. A couple of minutes later it had assembled a full landing page out of real components, populated with relevant, on-brand content. To make the point that a human stays in the loop, I dropped an image from the media library into the hero component and published. Then I switched to the Astro app, navigated to the same path, and there was the identical page — every decision the human and the model had made, rendered by the decoupled front end. A complete landing page, start to finish, in a couple of minutes.
The second beat pushed further. The marketing team wants a brand-new component: a call to action for a waitlist. I asked Canvas AI to build a full-width announcement banner with an announcement pill, a headline, a supporting line, and a primary call to action. After a short pause, the component appeared in the Canvas interface — colors on brand, formatting consistent with the rest of the library — with its code fully visible and editable and a live preview I could resize to check different breakpoints. I noted that in the real world you might refine the code yourself or ask Canvas AI to iterate, then saved it to the library, dragged it into the Executive Suites page, and published.
When I reloaded the Astro app, it threw a fatal error, exactly as I had planned. The layout now referenced a component the front end did not know about. One npx canvas push from the command line synced the components, a refresh brought the page back, and the new banner rendered cleanly in the Astro layout. That deliberate stumble made the architecture legible: content edits flow to the front end instantly, but new component code is a real, versioned artifact that moves through a real workflow.
I closed the demo by going back to the Context Control Center, because that is the intelligence that made the rest possible. Some context items are global and travel with every request. Others are scoped specifically to working in Canvas. Others still apply only to content about the Executive Suites program. All of them were assembled automatically behind those short prompts — which is why one sentence was enough to get on-brand, relevant output. I finished on the form for managing a single context item, showing the range of ways its use can be scoped and restricted. Compliance before generation, not review after.
Why Enterprises Can Trust It
For regulated and enterprise teams, governance is where this stops being a demo and starts being a decision. Drupal is model-agnostic by design: dozens of providers sit behind one abstraction layer, spanning cloud services like OpenAI, Anthropic, and Gemini as well as self-hosted options like Ollama and Mistral for data sovereignty. Swapping providers is a configuration change, not a rewrite of your schemas or your logic.
Agents act inside Drupal's existing permission model which includes the Access Policy API, so the access logic that already governs your people governs your agents too — no separate guardrail layer to maintain. Deterministic orchestration through the Event-Condition-Action (ECA) or FlowDrop frameworks handle rules-based logic that costs no tokens and never hallucinates, which is a useful reminder that the cheapest, most reliable AI call is often the one you do not make. Guardrails filter sensitive data before it leaves the server, and metering tracks token spend by user and role so finance can see what AI actually costs.
An Honest Read
It doesn't serve anyone to pretend one side wins everything, and I said so in Montréal. The headless-native platforms lead on real things: faster time to value, a cleaner developer experience, and more polished agentic tooling in market today. If those are your priorities right now, they are genuine strengths.
Where Drupal leads is open source with no lock-in and dozens of documented APIs, model-agnostic freedom, deep governance and orchestration, and fit for enterprise, multi-brand, and regulated environments. It is also worth remembering the shape of the thing behind it: an open ecosystem moves at the speed of everyone who needs it to, while a single-vendor roadmap moves at the speed of one company's priorities.
The Takeaway
The way I put it at the end of the talk: we gave up the editorial experience to go headless, and in 2026 we stopped having to. The original headless win is now additive with the editorial win, not traded against it. One structured content model can serve four consumers at once — a decoupled front end, editors in Canvas, internal AI agents, and the wider martech stack over MCP.
Drupal is not a CMS with AI features bolted on. It is a governed AI harness that happens to have been building the right foundations for 20 years. If you want to see it for yourself, start with Drupal CMS 2.0 and Canvas, then explore the AI, context, and MCP modules. For teams that would rather not set up and host Drupal themselves, Acquia Source CMS offers a fully managed on-ramp to the same platform. And if you are ready to help shape where this goes, the Drupal AI Initiative is where the work is happening.
