Terms from Home and The Loop, Zoomed In, in one table.
| Term | Plain-English definition |
|---|---|
| Model | The large language model itself. It only ever does one thing: given some text, predict what comes next. It has no memory of past requests unless that history is included in the current one. |
| Harness | The ordinary program wrapped around the model: it sends the requests, reads the model's replies, actually executes any tool the model asks for, and decides when to stop. The loop described on this site is the harness. |
| System prompt | A block of instructions sent at the start of every request, invisible to the end user, that tells the model who it is, how to behave, and which tools it has available. |
| Tool / tool call | A tool is a description — a name and a schema for its arguments — of something the harness can do on the model's behalf, like reading a file. A tool call is the model asking the harness to run one, by name, with arguments. The model never runs the tool itself. |
| Context window | The maximum amount of text (measured in tokens) the model can read in a single request. Since every turn re-sends the whole conversation, a long-running loop eventually has to trim, summarise, or clear that history to stay under the limit. |
| API request | One round trip to the model: the harness sends the system prompt plus the conversation so far, and gets back one reply (text, a tool call, or both). Every step of the loop that touches the model is a new request like this one. |
Still stuck?
Go back and re-read The Loop, Zoomed In with these definitions in hand, or see Web Ring & Links for where this page's understanding came from.