chesterdocs

#Configure an agent's model

An agent runs on an LLM adapter and environment from Systems, a model the provider lists, and a system prompt; temperature, maximum tokens and maximum steps limit how it answers.

The Configuration tab of the agent "Support assistant": (1) the System prompt editor, (2) the LLM provider card with no LLM adapter selected yet, (3) the Model settings, locked until an adapter and an environment are chosen, (4) the chat start path below, and (5) the Tools count in the header.

#In what order

flowchart LR
  A["1 · LLM provider<br/>adapter of a system"] --> E["2 · Environment<br/>with a credential"]
  E --> M["3 · Model<br/>from the provider's list"]
  M --> S["4 · Temperature, Max Tokens,<br/>Max Iter."]
  P["System prompt"] --> R["The agent's instructions"]

Open the agent, click Edit, and stay on the Configuration tab.

#Choose the LLM provider

  1. In LLM provider, click Select and pick an LLM adapter. Adapters that cannot call tools are greyed out — an agent needs tool calls. To create one, see Connect an LLM provider.
  2. Pick the Environment. It supplies the address and the API key; only active environments with a credential are offered. If just one qualifies, it is chosen for you.

Switching the adapter clears the environment and the model; Remove clears all three.

#Pick the model and its settings

Once adapter and environment are set, Model settings unlocks:

Setting What it does
Model Chosen from the models the provider lists. Reload models asks again. A model you had chosen is kept even if the provider no longer lists it
Temperature 0 (precise) to 2 (creative), 0.7 by default. Hidden when the adapter or the model does not accept one — reasoning models use the provider's default
Max Tokens The longest answer, 4096 by default
Max Iter. How many rounds of thinking and tool calls before the agent is stopped: 1 to 500, 50 by default

⚠️ Reaching Max Iter. ends the run without an error — the answer is whatever the agent had reached. A task with a defined response still has to deliver it; see below.

⚠️ The labels Temperature, Max Tokens, Max Iter. and the token count next to the system prompt are not translated, and numbers use German formatting (0,7).

❌ Provider-specific reasoning options cannot be set.

#Write the system prompt

System prompt holds the instructions the agent always follows — who it is, how it answers, what it must not do. The small counter estimates its length in tokens.

When a process asks the agent to do a task, the task's own Instruction for this task is added after the system prompt; it does not replace it. See Let an agent call its own tasks.

⚠️ In a chat the agent is told the current date and time; when a process runs it, it is not.

#A structured answer

A task can define the fields of its response on the Tasks tab. The agent then has to deliver them through a result tool, checked against that shape. If it finishes without doing so, it is asked once more; if it still does not, the step fails. Without a defined response, the caller gets the agent's text.

A chat and a start by message never use a defined response.

#What stops a deployment

An agent without an LLM adapter, an environment or a model cannot be deployed; neither can one whose environment is inactive or has no credential, or whose adapter cannot call tools. The messages are German.

⚠️ The API key is taken into the deployment. After changing the environment's credential, deploy the agent again.