New observations, the same weights
The system adds tool results to the context and asks the model for the next action. The trained weights stay fixed, and a longer conversation trains nothing by itself.
- Goal
- Action
- Observation
- Next action
Inside an AI agent.
Follow one task through an agent, from the goal to tools, approval and a clear ending. Turn the 3D model, open its stations and run a scripted agent step by step. It builds on the lesson inside a language model.
The agent, laid outLoading the 3D model
Drag to turn the model. Click it once, then scroll to zoom. On a phone, use two fingers. Click a station or its label to learn what it does. The amber parcels show the usual order of the loop as an illustration. The run below drives the coral highlight.
Run the loop yourself
A scripted agent, one step at a time
The decisions are written in advance and the tools use made-up data, so you can watch every step without a real model, account or network call. The station the run has reached lights up in coral on the model.
Goal
Your approval is needed
This action saves a made-up reply in the memory of this simulation. Nothing is sent.
0 of 5decision iterations
0tool executions
0observations in the context
Fixedmodel weights
What happened
A log of the scripted events. It shows actions and results and never a model’s private reasoning.
What the system knows
How the run ended
A common mix-up
Running an agent and training a model
An agent that works with more context can look as if it learns. Its model stays the same.
Examples change the weights
Training uses data and an objective to update the weights. It is a separate process. An agent can be connected to one, and its tool use alone retrains nothing.
- Examples
- Objective
- Weight update
Check yourself
Six short questions
Pick an answer to see a short explanation. Nothing is scored or saved.
Further reading
Where these ideas come from
The lesson is a teaching model. Real agents differ, and many use fewer or more parts.