How AI models work
You do not need a machine-learning degree to use 4RGE well. This short lesson explains what models are, how they behave, and how your organization controls which ones you can use — so the model picker and agent runs make sense.
What this page is for
This is a foundation lesson. When you are ready to run agents, start with Quickstart. For admin catalog setup, see Model control.
Who this page is for
The idea in one minute
A language model reads your prompt and predicts what text should come next — over and over until it finishes a reply. It learned patterns from huge amounts of text during training; it does not look things up in a database the way search does.
That last row is the important one for 4RGE: the model suggests; the agent loop reads files, runs tests, and asks you before risky steps.
Models in 4RGE
In the app, CLI, and cloud, a model is the AI engine behind an agent run. Your organization defines a catalog — which models and providers members may use. Admins set defaults and limits in Model control; members pick from what is allowed.
If the picker is empty or a model is missing, your org may not have enabled it — ask an admin, do not assume your account is broken.
Usage from every model call counts toward Usage under your org (and often a project). Heavier models generally consume more allowance than lighter ones.
How to pick a model
There is no single "best" model for everyone — task type, speed, and cost all matter. Use this as a starting guide; your org may only expose a subset.
Many developers prefer different models for the same repo — experiment within your allowed catalog, then stick with what matches your style. Admins can narrow the list so onboarding stays simple.
Switching models mid-task usually starts a fresh context — finish or discard the current run before changing.
Models vs agents
The agent is what makes model output useful and inspectable in your repository — not magic autopilot. See Agent overview for the full loop.
What to expect
Good habits:
- Ground the model — point at files, paste errors, describe constraints.
- Review diffs — treat agent output like a junior teammate's PR.
- One task at a time — smaller scope beats "rewrite the whole app."
Normal limitations:
- Wrong API names, paths, or config keys (hallucination) — verify with tools and tests.
- Context limits — very long chats drop older messages; start a new run if the agent loses the plot.
- Non-determinism — re-running the same prompt may wording-differ; check behavior, not exact text.
For deeper lessons: Hallucination & limitations, Context, Tokens & pricing.
Where to go next
Questions about models?
When models invent details — and how to guard against it.
How usage is measured and why model choice affects spend.
The plan → act → verify loop that wraps the model.
What admins configure — catalog, defaults, and blocks.
How 4RGE agents work in practice on desktop and cloud.