Agent configuration

This page provides a detailed reference for all agent configuration options. You can edit these settings at any time from the agent detail page in the dashboard.

Model selection

Only models that support tool use are available for agents. The model determines the agent’s reasoning capabilities and response quality. Select a model from the dropdown when creating or editing an agent.

Temperature

Temperature controls the randomness in the agent’s responses. Set this value based on your use case.

Value	Behaviour	Best for
0	Deterministic, precise	Code review, data analysis, compliance
0.5	Balanced	General-purpose assistants
1	Creative, varied	Content generation, brainstorming

Max tokens

Set the maximum number of tokens (1—200,000) in the agent’s response. Higher values allow longer responses but consume more credits. Leave empty to use the model’s default limit.

Note

The actual maximum depends on the selected model. If you set a value higher than the model supports, the model’s maximum will be used instead.

System prompt

The system prompt defines the agent’s personality, expertise, and boundaries. It is the single most important configuration option for shaping agent behaviour.

Best practices

• Be specific about role and expertise — tell the agent exactly what it is and what domain it operates in
• Define persona and tone of voice — specify whether the agent should be formal, casual, technical, or conversational
• Set clear boundaries — state what the agent should and should not do, including topics to avoid
• Include examples — provide sample input and expected output so the agent understands the format you want
• Reference available tools by name — mention the tools the agent has access to so it knows when to use them

Example system prompt

You are a senior code reviewer for a TypeScript project. Your role is to:
- Review pull requests for correctness, readability, and performance
- Flag potential security issues
- Suggest improvements with code examples
- Use the "search_codebase" tool to look up related code when needed

Be concise and constructive. Focus on actionable feedback.
Do not make changes directly — only suggest improvements.

Tools

Tools extend what an agent can do by connecting it to external functions. There are two types:

• Built-in tools — provided by Quant AI (e.g. web search, content retrieval). These are maintained by the platform and available to all agents.
• Custom tools — your own edge functions registered as callable tools. You define the function, input schema, and description.

Agents can only use tools you explicitly assign. Unassigned tools are not accessible to the agent, even if they exist in your project.

Caution

Removing a tool from an agent takes effect immediately. If the agent is mid-conversation and relies on a tool you remove, subsequent tool calls will fail.

Skills

Skills are reusable instruction sets that enhance agent capabilities. They are grouped by namespace when assigned to an agent. When you select a namespace, all skills in that namespace become available to the agent.

Skills allow you to share common instructions across multiple agents without duplicating system prompt content.

See Skills for details on creating and managing skills.

Vector database collections

Give agents access to knowledge bases for retrieval-augmented generation (RAG). The agent can search assigned collections during conversations to ground its responses in your data.

This is useful for:

• Internal documentation and knowledge bases
• Product catalogues and specifications
• Policy documents and compliance materials
• Any domain-specific content the agent should reference

See Vector Database for details on creating collections and ingesting documents.

Groups

Organise agents into groups for easier management and filtering in the dashboard.

Built-in groups:

• development — code review, debugging, technical assistance
• compliance — regulatory checks, policy review
• content — writing, editing, content generation
• analytics — data analysis, reporting, insights

You can also enter a custom group name when creating or editing an agent. Groups are free-form labels — type any name to create a new group.

Next steps

• Testing agents — test your agent configuration using the built-in chat interface
• Tools — create custom tools for your agents
• Skills — build reusable instruction sets
• Vector Database — set up knowledge bases for RAG