Agenten-Kompass
Guide overview

ModelPart 1 of 5 5 minutes

What does the model do inside an agent?

The model processes the current context and produces an answer or proposes a tool call. The runtime, the software around the model, executes permitted calls.

One model call, step by step

Shape cards enter a cube of connected nodes as input. A card with a proposal comes out, followed by a check mark.1Input2Model3Suggestion4validation
The four stations in the picture match the four steps. AI-generated illustration.

  1. Receive context

    The task, instructions and selected information form the input to the current model call.

    With LeaTask: a slot with Lea. Note from memory: Lea prefers mornings.

  2. Generate the next step

    The model generates an answer or a tool call with appropriate input values.

    With LeaThe model proposes searching the calendar for this week.

  3. Execute and observe

    The runtime executes permitted calls and returns the results to the model.

    With LeaThe runtime runs the search and places three free slots in context.

  4. Check and continue

    Feedback shows progress or errors. A further step or completion follows from it.

    With LeaThe slots are there, the invitation is still missing. So another model call follows.

TRY IT YOURSELF

What really changes the model?

Some things change the weights, the learned numbers inside the model. Others only supply working material for the current call. Which is it each time?

You upload your project documents to the chat.

Only the context The documents are working material for this conversation. The model does not permanently learn from them.

The provider trains a new model version.

The weights Training adjusts the weights. This is how learned patterns emerge.

You write: please answer in German.

Only the context Instructions in the prompt steer the output. Parameters stay unchanged during ordinary inference.

A team further trains the model on its own examples.

The weights Additional training can change the weights. That is a separate process outside the normal conversation.

TRUE OR FALSE?

Three things people often say.

The model runs tools itself.

False It generates a call. The runtime executes permitted calls and enforces access rules.

The same input always means the same answer.

False Outputs can vary for the same input. That is why checks need to fit the task.

Even a confidently worded result needs checking.

True Confident wording is no proof of success. Passed tests or traceable sources, for example, are meaningful.

Explore the mechanisms

Training shapes capabilities

Training adjusts weights: the numbers used by the neural network in its computations. This produces learned patterns. Further training can change those weights; supplying project documents in a conversation instead provides current working context.

Google: Training neural networksGoogle: Fine-tuning and prompt engineering
Inference uses what was learned

Using a trained model is called inference. It processes the current input using its existing weights. Examples and instructions in the prompt can guide the output while model parameters remain unchanged during ordinary inference.

Google: Fine-tuning and prompt engineering
The loop enables work steps

An agent connects repeated model calls with a runtime: the software controlling execution. Tool results add context. The model can then adapt its next step; the runtime enforces access rules and stopping conditions.

Anthropic: Building effective agentsAnthropic: Handle tool calls
Results need appropriate checks

Outputs can vary for the same input. Meaningful checks therefore address the actual task: passing tests, traceable sources or satisfied requirements. Confident wording alone is unsuitable as evidence of success.

OpenAI: Evaluation best practices

Fundamentals and primary sources reviewed on 5 November 2026. The examples explain the principle; specific products may differ.