We made WordPress optimise its own pages.
It can go on autopilot, 24 hours a day, 7 days a week. The video shows it running (watch the video here).
Here’s what we actually did:
The premise is simple. You point an agent at the blocks that matter and let it test against itself:
- Identify the high-value blocks based on traffic and engagement (the hero, campaign promos, etc.).
- Connect Claude Code or Codex to those blocks and your analytics.
- Have it run its own tests with self-generated variants, using your theme and styles as context.
- It learns from each run and creates increasingly better variants on each turn.
In other words, it’s self-optimising towards higher conversion, even while you’re sleeping. Under the hood, a bandit shifts traffic toward whatever’s working instead of splitting it 50/50, and a Bayesian model only calls a winner at 95% confidence. There’s no huge architecture, but rather good prompting, an aggressive evaluation model, and everything thrown in a loop.
This is only possible because of WordPress’ open-source AI foundations. In 2026, core shipped the plumbing that lets an agent safely operate a site — the Abilities API and the MCP Adapter — so 40% of the web quietly became agent-ready. In our “self-optimising” case, we’re using Accelerate’s AI toolkit, which can remotely control and test a WordPress setup. We’re early, but we assume we’ll see more projects like ours come out this year.
Below, a screenshot of the setup running through such test-sets and improving the block until it hits a natural conversion ceiling.
Getting this right involves a lot of pieces; mixing conservative brand guidelines with chaotic randomness, making significant changes to the blocks yet still fitting the current style and layout.
Then it has to be relentless about iterating on winners, so each run chases a genuine new high rather than a local maximum. Copy-only changes will only bring you so far, but with this setup, you’re altering imagery, layout and more.
We’re still testing and iterating how this all comes together on Accelerate, but essentially we set a hidden “winning” benchmark that Claude or Codex couldn’t see, and let it relentlessly pursue higher numbers against it. It learned from the top performers and dropped losers early, only to replace them with drastic new variants that brought fresh ideas into the next test run.
This path follows nature, where evolution and extinction happen at breakneck speeds, starting and ending experiments while you sleep. Where it isn’t pure A/B testing you can apply the same to personalisation, covering more audiences with bespoke messages. Put a human in the loop wherever you’re most comfortable.
Running websites, and the conversion and profit that come off them, is changing at speeds we’ve never seen before. Entire businesses optimising by hand will get eclipsed by the ones testing against themselves overnight.