Glossary A–Z
Define the word before you argue about the system
About ten sample terms ship with v1. Add a Markdown file in content/glossary to publish a new entry.
A
B
Sample content · easy to replace
Benchmark
A frozen set of tasks plus a scoring rule. It is not the same as your production job.
C
Sample content · easy to replace
Context window
How much text a model can see at once — instructions, history, tool results, and retrieved notes.
H
Sample content · easy to replace
Harness
The runtime around a model — loop, tools, memory, permissions, and stop conditions.
L
Sample content · easy to replace
Language model (LLM)
A model trained to predict text; today's "AI agents" almost always wrap one.
M
Sample content · easy to replace
Multi-agent
More than one model (or role) collaborating, often as planner, worker, or critic.
P
Sample content · easy to replace
Prompt
The instructions and context given to a language model for a single step or a whole run.
R
Sample content · easy to replace
RAG (retrieval-augmented generation)
Fetching documents or traces at run time and stuffing them into the model context.
T
Sample content · easy to replace
Tool use
Letting a model call APIs, a browser, a shell, or other programs, then reading the result.