In short
The intelligence an agent calls when it needs to read, write, or decide. Replaceable. When a better one ships, the same instructions run on it and the results improve.
A model is the trained system that turns text in into text out: reads a draft and structures it, reads a lead and writes a brief, reads a transcript and lists the decisions. Several labs ship them, they improve every few months, and they are priced per use.
We treat the model as a component. The agent's instructions say what to do; the model is what does the thinking on each run. We pick the best model for the job and change it when a better one appears. The instructions do not change.
This is the mechanism behind the promise that what we build gets better over time. The part you own is the part that persists. The part that improves is swapped underneath it.
Example
What it looks like.
The same instruction file ran on one model in spring and a stronger one in summer. The summer posts had fewer structural corrections and the validator failed less. No rule was rewritten.Related
Read next.
- inferenceRunning a model to get an answer. It costs money per call in the cloud, or electricity on your own machine. Where it runs is a choice.Read the definition
- harnessThe software that runs an agent: loads the instructions, connects the tools, keeps the schedule, delivers messages. Replaceable. The instructions are not.Read the definition
- context windowEverything the model can see on one run: the instructions, the files it was given, the conversation so far. Outside the window, it does not exist.Read the definition
- large language modelA model trained on enormous amounts of text that reads and writes language. The engine inside every current agent.Read the definition