The working vocabulary of building software with AI agents — the words, and the ideas under them.
When people and AI agents build software together, a shared vocabulary piles up fast: handoff, ladder, probe, vacuous pass, drift. Most of it gets absorbed without anyone ever saying what it means. This is that vocabulary written down, so it can be learned on purpose and taught.
Each word is a short handle for a bigger idea. That compression is the point: for an agent a short word is less context to carry, and for a person it is one thing to hold instead of a sentence to re-read. Much of it is borrowed from project management, engineering and law; some of it grew out of how agents work. Every entry says where the word comes from, who named it first, and where our use differs.
Alongside our own working vocabulary there is a section for the wider field: the words in AI news, papers, product docs and online debate — model, token, agent, hallucination, AI slop, p(doom) — explained plainly and linked to our sharper in-house terms where they meet.
The terms are sorted by who already knows them. Colours are the same everywhere on the site.
A handful of terms that come up constantly. Click one to open it with its full entry and its neighbours.
The ideas with the least prior art outside our own work — judged from each entry's prior-art notes, then checked against how often the phrase appears on Wikipedia and in Google Books. Treat "coined here" as "we found no earlier name", not a claim of priority.
The terms that come up most across our docs, commit messages and working sessions — counted by a script, with ordinary-English matches discounted.
Every term has the same parts, so you can skim for the one you need.