Core Concepts

An agent is defined by its configuration. No code required.

System Prompt

The system prompt is the agent's instruction manual. It tells the LLM who it is, how to behave, what to do, and what not to do. The system prompt also instructs the agent on how and when to use its tools.

Tools

Tools are external services the agent can call, provided as MCP (Model Context Protocol) servers. You provide the MCP server URL when configuring the agent. Each tool is self-describing — it tells the LLM what it does and what parameters it accepts. The agent's LLM decides when to call a tool based on the conversation and system prompt.

If your MCP server requires authentication, you provide secrets alongside the URL as literal header values. The platform stores them with the agent's configuration, masks them in API responses, and passes them as-is to the agent's runtime.

Sub-Agents

An agent can delegate tasks to other agents. The parent agent's LLM decides when to call a sub-agent and incorporates the result. Sub-agent relationships are directional by design — A can call B, but B cannot call A unless explicitly configured. Sub-agents must exist and be owned by the same account; the platform does not detect or reject cyclic or self-referencing sub-agent chains (including via update), so avoid creating them yourself.

Sessions

A session is a conversation thread. Messages in the same session share context. You can have multiple sessions with the same agent. Sessions are created automatically when you chat.

Model

The LLM that powers the agent. You select the model when creating an agent.