A recent piece argues most AI agents fail because teams build them top-down and hope the model holds it together. I agree, but WordPress has no long-running PHP process, so that orchestration has to live in hooks, the database and a queue.
Every AI plugin ends up needing a threshold for when to auto-apply an answer and when to hold it for review. The confidence number most of us gate on tracks how fluently the model produced its tokens, which is a different thing from whether the answer is true.
A Towards Data Science tutorial calls a small Python chat loop an AI agent. The client setup is worth keeping, but it has no memory, no tools and no stop condition, and a WordPress build has to add all three.
A Towards Data Science piece builds one retrieval system four times, from TF-IDF to fine-tuned DistilBERT, and the fine-tuned model flips its answer between two epoch counts. For a WooCommerce store, that argues for pretrained embeddings and no training.
A store’s order-value histogram can show two peaks at thirty days and one hump at a year, because the bin count never changed while the data did. Here’s how I pick bins server-side, cache the edges, and stop the guessing.
On client sites, agents write plausible code against a stack that is not there. Specs help when they record what your WordPress install actually runs. These are the facts I pin down first.
WordPress Guidelines keeps agent memory, skills and instructions in an ordinary custom post type, and it’s headed for Core. Here’s what that means for plugins with their own memory stores, which permissions to test first, and when a separate table still makes sense.
Angle encoding costs you qubits and amplitude encoding costs you gates, and both get paid when data crosses into the system. WordPress has the same problem with autoloaded options and fat transients, and that one is cheaper to fix.
I read the WordCamp Europe 2026 schedule as a prep list. Register one Abilities API entry on staging, run the WordPress 7.0 upgrade on a copy, and pick the sessions you will actually use before Kraków or the livestream.
AWS shipped the Agent Toolkit, which gives coding agents skills, live docs and an MCP server. The part I care about is the guardrails: IAM condition keys that mark agent calls, and CloudTrail records of what the agent did. Set those up before an agent gets near a client account.