The AI glossary for people who ship software
Most AI glossaries are written for investors. This one is written for the person who has to make the thing work, so every entry answers the question you actually had: what is this, what is it not, and what changes in my code when I use it. Definitions come from primary sources, each page carries the date it was last checked, and where the industry uses a word two different ways, both are named.
If you are reading one entry, read one of these first:
- Confused about what an agent is → agent: the word describes a loop, not a product.
- Wondering where state lives → harness: the model keeps nothing, the harness resends everything.
- Hitting limits on long tasks → context window and memory are two different problems.
- Making a model do something → tool calling: the model asks, your code acts.
Every term
5 AI terms, defined in one paragraph each
Each card gives you the whole definition in one sentence. Open the page when you need the mechanics, the worked example and the sources.
Explainers
How the moving parts actually work
Some questions do not fit in a definition. These walk through one mechanism end to end, with the failure modes named.
Common questions
Common questions about AI terminology
Why do AI terms mean different things in different docs?
Because most of this vocabulary was coined by product teams, not by a standards body. Words like agent, memory and reasoning were marketing labels before they were technical ones, and each vendor anchored them to its own implementation. Where a term genuinely has two meanings, this glossary names both and says which one a given tool uses, rather than picking the definition that sounds tidiest.
Is this glossary about a specific model or framework?
No. Definitions are written to hold across providers, and where behaviour differs the entry says so and links the vendor's own documentation. That matters most for tool calling and context windows, where the shape of the API and the size of the limit are decided per model, not per concept.
How current is it?
Every page carries the date it was last checked, and the checks are scheduled rather than occasional. Terms whose facts move (limits, pricing, API shapes) are re-read more often than terms whose facts do not (definitions of concepts). If an entry has drifted, the date will tell you before the text does.
Do I need to understand the maths to use this?
No. Nothing here requires linear algebra. The entries describe what a mechanism does and what it costs you as an engineer: what gets sent over the wire, what the model can and cannot see, and where a design falls over. Where a paper is the primary source it is linked, so you can go deeper on the entries where it helps.
Why does a CI/CD company publish an AI glossary?
Because the two subjects stopped being separate. Agents run in pipelines, pipelines run models, and the operational problems (isolated environments, credentials, reproducible runs, logs you can read afterwards) are the same problems continuous delivery has been solving for years. Where a term touches that overlap the entry says so plainly, without turning the definition into a pitch.