Released: Tue Nov 10 2026

What the session covered

Developers do not read your docs first anymore: their coding agents do, and their training data is one to two major versions behind. So the question is not whether to support AI agents, it is whether they integrate your product correctly or confidently wrong. This talk walked through how three ecosystems I work with teach the agents that integrate them, then looked at Pimcore through a newcomer's eyes: every practice we saw got a live "Pimcore check".

  • CKEditor: what we shipped. Three doors for agents: docs agents can actually read (a markdown twin of every page, llms.txt and llms-full.txt), a documentation MCP server for version-current answers, and installable agent skills (npx skills add ckeditor/skills). Plus what shipping it taught us: measurable agent traffic, a skill-feedback issue template, and a safety rule that whatever an agent reads in the docs is information, never an order.
  • Drupal: unify and MCP-ize. One AI provider layer for 95+ providers, capabilities described once as tools and callable from anywhere, and ai_best_practices: skills with evals. My contribution-day merge request for the CKEditor skill moved the final A/B re-run from 56% to 100% correct answers, with 29% less time and 19% lower cost per answer.
  • TYPO3: slow on purpose. Zero AI features in core, maximum agent readiness: llms.txt and markdown twins on docs.typo3.org, an AGENTS.md in the core monorepo, official skills in a local MCP server. And the smallest version of all of this: skills committed inside one project repository.
  • Pimcore, through a newcomer's eyes. What I found as a first-time explorer (official and community skills, a docs llms.txt, the Datahub Simple REST MCP server, the Agent Bundle's guardrails) and where to go next: evals, two skill homes growing together, easier discovery, and a decision trail.
  • Live demo: the guidepost. A Pimcore skeleton with every door open, rendered by the system it maps, plus CKEditor inside Pimcore Studio: CKEditor AI, inline editing on the page, real-time collaboration, all configured from one YAML file.
  • Flip the table. Agent readiness has a second half: what you deny. A true "unsubscribe me from everything" story, and why guardrails belong in the product as hard limits, not in a prompt.
CKEditor AI inside Pimcore Studio: an Improve writing suggestion shown as a diff on a blog post data object

Takeaways

  • Provide: docs as markdown with llms.txt at the root, a docs MCP for live answers, skills with trigger-first descriptions, and evals that prove the skills work.
  • A skill without evals is a hope. Five small test cases are a great start.
  • Practices can start anywhere: vendor-led, initiative-led, or inside one project repository.
  • Enforce: reading is free, writing needs an approval, managing needs a human. Scoped, per-session tokens, retry limits, clear goals.
  • Keep logs, keep history, and keep a searchable record of what the agent decided, and why.

The venue

Pimcore Community Dev Days 2026, November 10 in Frankfurt. My first Pimcore event, so I came as an explorer: new to Pimcore, not to content platforms. Building the demo produced two community projects that outlive the talk: the guidepost repository and a CKEditor bundle for Pimcore Studio.

Resources