Vectorless RAG is accurate, slow, and expensive to index. Proxy-Pointer RAG puts the document structure into the embeddings and metadata instead: regex skeleton trees, breadcrumbs on every chunk, and pointers back to the full section.
Claude Code will scaffold a WordPress plugin faster than you can spec one, which is the problem. The workflow I use: a CLAUDE.md spec first, the terminal agent second, and a manual review of everything before it ships.
Two bad defaults for agent UIs: hide everything behind a spinner, or stream every raw log. What I map instead, and how I decide which agent decisions a user actually needs to see.
The Core AI team has a draft guide for anyone presenting on Abilities, the AI Client or the MCP Adapter at a WordCamp or meetup. Comments on it close on 21 April 2026.
Embedding models turn text into coordinates, and that is what vector search is really matching on. This covers tokenization, chunking, cosine similarity and fine-tuning a model so it fits your own catalog rather than the general web.
AI tools guess at your store because they cannot see it. MCP gives them real context: your files, your data, your WP-CLI. We talk it through at WooCommerce office hours on April 15, 2026, on Slack.
Pinecone and Chroma are a lot of machinery for a few hundred Markdown notes. The Google Memory Agent Pattern keeps structured memories in SQLite, consolidates them in the background, and loads them straight into a large context window.
Most people run one Claude Code agent and then wait for it. Git worktrees let you keep several going in separate folders on the same repo. How I lay that out in Warp or Tmux, plus the shared local database problem nobody warns you about.
The AI Client and Connectors API change how plugins talk to an LLM in WordPress 7.0. Notes from the Dev Chat agenda, plus the PHP refactor from a hardcoded wp_remote_post call to a registered ability.
MCP lets an AI pull data out of Google Drive, read old repos and load BigQuery instead of just suggesting Python. Includes the 150GB log file one unattended loop produced, and why domain knowledge still decides the answer.