Configuring an agent: system prompt and model
The agent configurator in depth: system prompt, model choice, parameters, and an advanced mode for power users.
The configurator is where an agent gets its character. You set the system prompt, model, and parameters here. There's a simple mode (the key settings) and an advanced one for fine control.
System prompt
The system prompt is the single most influential setting. A few principles for a good one:
- Role in one line: "You are a financial advisor for self-employed in the US."
- Output format: "Reply briefly, in bullets, with numbers."
- Boundaries: what the agent must not do ("don't give legal guarantees").
- Context: stable facts — audience, tone, language.
Prompt edits usually change the result more than switching models.
Choosing a model
| Model | When to pick |
|---|---|
deepseek-v4-pro | Top quality and reasoning — the default for new agents |
deepseek-v4-flash | Fast everyday tasks, dialogue |
glm-5.2 | Strong general-purpose alternative |
All of these are available over /v1/chat/completions; the DeepSeek models are also served on the
Anthropic-compatible /anthropic/v1/messages endpoint. The full list is available via the
developer API.
Parameters
- Temperature — output creativity. Lower is steadier and drier, higher is more varied.
- Max tokens — reply length cap. Affects token spend.
- File operations flag — allows file operations where applicable.
Advanced mode
Advanced mode exposes a direct editor for the system prompt and all parameters. It's handy when you need precise control or are porting a ready prompt from another tool.
How to test changes
- Save the agent.
- Open the chat and ask a typical question.
- If the result isn't right — refine the prompt and repeat.
Prompt iterations usually beat model switches. Token spend on tests shows up in Usage.