Most agentic CMSs draft, translate, tag, check the brand voice and publish. But what context and data outside the project can the agent access? That list is short.
Give an agent no context, and it writes the average based on training data. Give it access to the product specifications and the competitor landscape, and suddenly it starts to write posts that differentiate your business.
Every vendor agrees on the problem
Read the pages that rank for "agentic CMS" and one idea comes up again and again. The agent needs context.
Hygraph says it directly: "Traditional CMS platforms put the entire burden of context on the user." Frase makes context its first criterion, because "a system that never saw the research cannot act on it later." Amplience asks if a CMS can pull in context "so agents aren't generating copy in a vacuum", and warns that without it, AI tools "will simply hallucinate faster."
They are right. Then each of them solves it the same way: the context has to be inside their platform.
What is an agentic CMS?
An agentic CMS is a content management system where AI agents plan, research, write and update content from the real context of your business, and a person approves what goes live.
Most definitions stop at the first half. Optimizely defines it as a CMS that "uses autonomous AI agents to plan, generate, personalize, distribute, and continuously optimize content." That tells you what the agent does. It does not tell you what the agent knows. The second half is where most products fall short.
What each agent works on, in the vendors' own words
| Vendor | What its agent works on | Built for |
|---|---|---|
| Hygraph | Content "with full awareness of schema, validations, and publishing rules" | "enterprises with strict compliance, content standards, and review processes" |
| Optimizely | "Agents act on data directly within the CMS" | Enterprise marketing, e-commerce and B2B teams (its listed use cases) |
| Amplience | Context from your PIM, DAM and commerce platform, such as "inventory, pricing" | Enterprise retailers |
| Kontent.ai | Content checked "against company tone and terminology standards" | Mid-sized and large organisations (CMSWire) |
| Frase | The research "a page was built from", plus rankings and traffic | Content marketers, SEO teams and agencies |
| Cosmic | CMS content, plus "Slack, GitHub, the browser, external APIs" | Teams on a headless CMS |
Some of these agents reach far. Amplience connects to product databases. Cosmic agents can work in Slack and GitHub. That is real reach. But look at what it reaches: other business systems, each with an API, a login and often a contract. That is the context of a large company.
Small & mid-sized company data lives in folders
A company of three people does not have large enterprise systems. Its context sits in places like these:
- A folder of product notes and client briefs on someone's Mac.
- A Google Drive or Dropbox folder with brand files and old proposals.
- Release notes or a changelog.
- Search Console, with the queries the site already shows up for.
- The competitor sites open in browser tabs.
- The heads of the people who work there.
Janus Boye, an industry analyst, puts the bigger point well: "content management systems are actually rarely used to manage content." The management, he writes, "is often orchestrated using email, Excel and Word." The real context of a website was never inside the CMS.
A hosted agent only sees the files on your computer if you upload them or paste them into a chat. Every upload is a copy. Copies go out of date. The burden of context is still on you. It just moved from the editor to the upload button.
Agentic CMS, headless CMS or an AI button?
Three terms get mixed up.
A headless CMS is about where content is shown. It keeps content apart from the design and sends it to a site or app through an API. Frase puts it simply: the two "answer different questions, and you can have both at once."
An AI button is a writing tool inside the editor: Generate, Rewrite, Translate. BCMS lists it under what an agentic CMS is not: "An AI writing assistant." Frase calls it "a faster pen."
An agentic CMS takes a goal and does the steps itself. Cosmic draws the line like this: "An AI feature makes an individual editor faster. An AI agent changes how many editors you need." For a small studio, that means something simpler. The content work that never got done now gets done.
Cosmic also admits that "most vendors claiming 'agentic' are at Level 1 or Level 2", which in its scale means AI features and copilots. The word is ahead of the products.
How Splitleaf handles context
Splitleaf is a desktop app for Mac and Windows. Your site is a folder of files on your computer, and the agent works where your context already is.
- Your site. The agent reads every page and post, and keeps a short profile of your product, your audience and your voice. You never fill in a form.
- Your folders. Link a folder in the Assets panel: product docs, release notes, or a Drive or Dropbox folder that syncs to your computer. The agent reads it where it is. Splitleaf never writes to, renames or deletes files in a linked folder. No upload, and no copy that goes stale.
- Search Console. Connect it once in Site settings. The agent reads clicks, impressions and position for each page. It can only read.
- Your competitors. The agent reads their blogs, sitemaps and feeds, and finds the topics they cover that you do not.
- Your rules. You write rules for the site and for each collection, and the agent follows them every time.
The agent stages every change. You keep it or undo it. It never publishes. It runs on your own key from Anthropic, OpenAI or OpenRouter, and it also works with Claude Code.
Data ownership note: When the agent reads a file, the text goes to the AI provider you chose. The file itself stays on your computer, and Splitleaf never sees your local files or site folders.
This post is an example
The agent planned this post the same way. First it checked Search Console. Splitleaf.eu had zero impressions for any search with "agentic" in it between April and October. Then it read the pages that rank for the term, the ones quoted above, and saw the gap: all of them are hosted platforms, and most are built for large companies. It wrote a brief with sources, search terms and a plan. It presented three angles for the post, and wrote the draft for the selected angle.
See how it works on the agentic CMS page.
Evaluating an agentic CMS
- What can your agent read? Ask for a list. Then ask which items on that list live outside your platform.
- How does my context get in? By upload, by paste, by a connector, or does the agent read files where they already are? Every copy is one more thing to keep up to date.
- Can it show the source for each claim? An agent that writes without sources writes the average, or makes things up.
- Does anything go live without me? Ask to see where you approve a change, and how you undo one.
- Whose agent is it? Can you choose the AI model? Can an agent you already use, like Claude Code, work on your content? Frase is blunt about this: an agentic CMS that only its own vendor's agents can operate "is a walled garden with a fashionable name."