Skip to content
All writing
/uncategorized 5 min read

The new WordPress era: from developer to architect

What I showed at WordCamp Portugal 2026 about agentic engineering for WordPress. The 4 D’s, the kit, the MCPs, and the real cases running in production.

The new WordPress era: from developer to architect

At WordCamp Portugal 2026 I talked about what has shifted in my work since December 2025: I stopped writing code and started doing agentic engineering, designing the system that writes the code for me. This post recaps the talk, puts all the material in one place, and points to the kit and the open source projects you should take home.

If you got here from the QR code on the last slide, welcome. If you’re trying to figure out why so many seniors are suddenly doing more with fewer clicks, the summary is below.

Table of contents

The 4 D’s

Anthropic’s AI Fluency Framework proposes four competencies for working with AI effectively, efficiently, ethically, and safely: Delegation, Description, Discernment, Diligence. In the talk I walked through them in a different order (Delegation, Discernment, Description, Diligence) because that’s the order that makes sense in the flow: I can only describe what the agent should do once I know how I’m going to evaluate the output. Criteria come before the brief.

Delegation is deciding what you do yourself and what you hand to the agent. These days I delegate implementation, scaffolding, and routine refactoring. I keep architecture, security, and product decisions for myself. If it’s irreversible, it’s on me.

Discernment is the line between vibe coding and agentic engineering. Without tests, it’s vibe coding with extra steps. With tests, the agent has objective feedback to know when to stop. Tests are literally the dividing line between the two modes.

Description is the brief you hand the agent: CLAUDE.md, AGENTS.md, skills, prompts, MCPs. It’s the set of persistent instructions that keeps the senior dev who just joined the team from interrupting you every five minutes. The agent is exactly that new team member.

Diligence is what ships with your name on it. Pre-commit hooks that refuse code without phpcs. Read-only sub-agents auditing before any merge. The agent isn’t accountable; you are. This is where the professional and the amateur part ways, and where the senior engineer’s job changes for good.

What I showed live

The talk itself was pure narrative, no live demos, so the 55 minutes could carry the argument. But the artifact exists and lives in the repo: WC Faker AI, a WooCommerce plugin that calls wp_ai_client_prompt() (the new native WP 7.0 API) to generate 5 products in a single click from Tools > WC Faker. Everything I argue in the talk is mirrored in four commands inside demo-plugin/ that walk the full cycle:

  • /talk:01. Plugin bootstrap with the agent explicitly citing the conventions from CLAUDE.md (slug, namespace, prefixes) in the docblock. Description in action.
  • /talk:02. The wordpress-development skill scaffolding the admin page, REST, and wrapper around wp_ai_client_prompt(), with the wp-devdocs MCP confirming signatures instead of guessing. The AI call only fires when the button is clicked.
  • /talk:03. Read-only security-reviewer sub-agent on the diff. It lists issues. It doesn’t touch them.
  • /talk:04. git commit with a pre-commit hook running phpcs. Refused. CLAUDE.md is advice; hooks are law.

Clone the repo, open demo-plugin/, and the four commands live under .claude/commands/talk/. Run them in order and you get the full cycle: brief with conventions, generation with the right tools, review without trust, and the hook closing the gate.

Real cases of agentic engineering

The kit is not theory. It is four production sites built with exactly this workflow.

aitradedoctor.com is a SaaS for analyzing trading scripts. WordPress plus a custom plugin, OpenAI gpt-4-turbo in JSON mode, Stripe with three plans, MagicLink, AffiliateWP, AES-256 on the owner gateway so master prompts stay blind to developers. Built end-to-end with this workflow, live since March.

veritaslandco.com is a land brokerage in Texas. WP Engine, custom theme, Landio-style half-map search, Whitetail Properties-style video hero, MLS import engine. Four phases, surgical delivery.

winixamerica.com is the case I picked to make Diligence concrete. WooCommerce, Braintree, a custom plugin (pl-winix-helper) with a region pricing engine for the US and Canada, telemetry across 8 checkpoints, AvaTax sync. In April the checkout was failing for Canadians: the gateway agreed with the checkout in USD but the region was CA. One week of investigation, eight telemetry checkpoints, three deploys, daily emails to the client. The agent did most of the rounds. I decided when to go to production. Always me.

susanaschmitz.com is my wife’s site. She’s a photographer in Matosinhos. Headless architecture: WordPress as backend, Astro as frontend. The stack decisions were discussed with the agent; what shipped is two applications serving the same content in two different ways.

Take it home

Three open source packages, all from Pluginslab.

The kit: wp-agentic-kit. A complete starter for any WordPress project. Ships with annotated CLAUDE.md and AGENTS.md, a settings.json with a WordPress-focused allowlist, a .mcp.json with five MCPs ready to go, the wordpress-development skill with security and block references, three sub-agents (security-reviewer, gutenberg-block-builder, rest-endpoint-builder), and post-edit plus pre-commit hooks. Includes setup.sh to customize everything in one go. It’s exactly the set I use on my own projects.

The MCP servers:

  • wp-devdocs-mcp. The canonical reference for WordPress functions and hooks. Your agent stops guessing.
  • wp-blockmarkup-mcp. Schemas and validation for Gutenberg block markup. Useful for tests, migrations, and content generation.
  • wp-playground-mcp. Ephemeral WordPress instances to test plugins in seconds.

What’s next: wp-agentic-admin. A plugin that drops an AI Site Reliability Engineer inside wp-admin. WebLLM in the browser (Qwen via WebGPU), 100% local, GDPR by default. Fourteen abilities, four workflows, a ReAct loop that picks tools as it observes. v0.10 shipped at the CloudFest Hackathon; v0.11 is in its final sprint right now. If you haven’t seen what “WordPress that diagnoses itself” looks like, watch v0.10 and keep an eye on what’s coming next.

Where I am in May 2026

Back in January I wrote that WordPress developers had to adopt this model now. A few months in, what’s changed is the confidence. Every Pluginslab project runs through this workflow today because it delivers. Clients get features that would otherwise take weeks, and I sleep knowing the hooks won’t let me ship something dumb.

If any of this makes sense to you, clone the kit and use it. If something looks wrong, talk to me at hi@pluginslab.com. The slides live in the talk’s private org repo; the useful material is all in the links above.

Marcel Schmitz. Founder of Pluginslab. More about what I do at pluginslab.com.

Related: The Senior Engineer’s Job Just Changed, the broader argument about the senior engineer’s job in 2026.