Agent Configuration
An agent has these configurable fields:
| Field | Required | Description |
|---|---|---|
| name | Yes | 1-64 characters. Must start and end with a letter or number. Unique per account. |
| description | No | Summary of what the agent does (max 512 chars). Shown in the agent list. |
| model_provider | No | The model provider. Currently only “anthropic”. Default: “anthropic”. |
| model | No | The LLM to use. Currently available: “claude-opus-4-6” (default). |
| system_prompt | No | Instructions that define behavior (max 8,192 chars). If empty, the agent uses “You are a helpful assistant.” |
| tools | No | MCP server URLs the agent can call, with optional authentication secrets. |
| sub_agents | No | IDs of other agents this agent can delegate to. Must be your own agents. |
System Prompt Tips
The system prompt defines everything about the agent's behavior. MCP tools are self-describing (they include their own parameter schemas), so you don't need to document tool APIs — just tell the agent WHEN and WHY to use each tool.
- Persona: “You are a senior DevOps engineer specializing in Kubernetes and AWS.”
- Constraints: “Only answer questions about infrastructure. Decline unrelated requests.”
- Tool usage: “When asked to scan code, use the vulnerability scanner tool.”
- Output format: “Always respond with a brief explanation followed by a code block.”