From manual coordination to an agentic CMS: how we built our content system with Claude Code

At Diverger we produce content continuously, in several languages and for different communication channels, aimed at very diverse audiences. The problem was not a lack of content, but that it was badly scattered: some lived in the content management system (CMS), some in loose documents, some in translation tools and some in internal conversations. Nothing connected any of those pieces to each other.

That disconnection fell on the Communications team, which acted as the integration layer between all the tools. Their work consisted, above all, of recovering what had been published before, copying the necessary information from one place to another, adapting the message to each channel, preparing translations and coordinating review across several people on the team. It was a highly manual, repetitive task that often ended up taking longer than writing the content itself.

A language model like Claude can write, classify or summarise text without trouble, but that on its own did not solve our situation. The model by itself does not know who has permission to publish what, which channels exist, which languages are active or what Diverger's brand rules are.

We needed something broader: a well-defined system around the model that set out clearly what information each AI agent may consult, what tools it may use, what actions it may carry out on its own, which it must limit itself to proposing so that a person approves them, and how everything the system did is recorded afterwards. We call that set of rules and structure an agentic harness.

Out of that need came the idea of building an agentic CMS: a system of AI agents that, following rules defined by Diverger, would prepare content for multiple channels and languages, while the Communications team always kept the final say over what was published.

The system we built was organised in five layers, each with one clear responsibility. The first was a sensing layer, which watched the state of the CMS and the connected tools at all times without modifying anything, since its main job was to report what was going on. The second was the context layer, which combined intelligent search over existing content with fixed rules capturing Diverger's brand voice and editorial policies, so that any agent knew how we spoke and what was and was not allowed. The third was the orchestrator, which split the work between specialised agents (one focused on writing, another on translation), just as a team coordinator would. The fourth layer was the execution tools: the concrete actions the agents could carry out, each with a clear contract covering what it did, what permissions it required and what limits it had. No agent had free access to the CMS; they could only use the tools that had been explicitly assigned to them. The fifth and last was the governance layer, which applied company policy, required human approval where appropriate and recorded every action taken, so that it was always possible to reconstruct what happened and why.

Claude Code let us build this system in a reasonable time, mainly for two reasons. The first was that we used it as a development environment, letting us work with the whole code repository and its documentation as context in every working session. That was key because each new capability we added affected several of the five layers at once, and we needed that overall view. The second was that we used it as the execution environment for the agents themselves: its way of working with bounded tools, where each agent has an explicit list of what it may and may not do, with no free access to the CMS, fitted naturally with the control and governance architecture we wanted to build.

Today we have an agentic CMS in production, working on the real content of the Diverger brand, coordinating writing, translation, search optimisation and channel preparation tasks, all from a single workflow. Bringing in Claude Code was not only a way to save time; its greatest value was letting us separate two things that had previously been entangled. On one side, the technical infrastructure that ran and controlled the agents, and on the other, Diverger's own knowledge of its brand, its editorial rules, its permissions and its approval processes. Thanks to that separation, we have an agentic CMS that automates the repetitive work of coordinating content across channels and languages without giving up editorial control, since that final decision remains, as always, the responsibility of the Communications team.