Skip to main content
An agent is assembled from parts: a playbook and the tools it can call, a function with its branches and variables, an API tool, an MCP server, the project’s variables, and the documents its knowledge base retrieves from. All of them can be created and changed from outside the Studio, which is what makes generating agents from a template, importing them from another system, or keeping a set of projects consistent with each other possible.

Endpoints

Agent

Playbook

Function

Function path

Function variable

Tool

API tool

API tool variable

Variable

MCP server

MCP tool

Knowledge base document

How the build endpoints are shaped

Every resource follows the same six-verb pattern, so learning one teaches the rest: The batch variant matters when you are importing. Creating fifty variables one call at a time is fifty round trips and fifty chances to end up half migrated; one batch call is atomic from your side.

Scoping every call

Build endpoints act inside one environment of one project, so both are required query parameters:
Editing main changes what your team sees in the Creator. To stage changes without touching it, clone an environment first, build against the clone, and merge when you are satisfied.

How the pieces relate

  • A playbook is the unit of agent behaviour. It holds instructions and the tools it may call.
  • Tools are what a playbook can do. An API tool calls an HTTP endpoint; a function runs your own code; an MCP server contributes a set of MCP tools discovered from that server.
  • Function paths are the branches a function can exit through, and function variables are its inputs and outputs. Both belong to a function and are managed separately, so you can add a branch without rewriting the function.
  • Variables are project state that persists across a conversation.
  • Knowledge base documents are the corpus the agent retrieves from. Create them from a URL, from text, or by uploading a table.

Where to go next

Once an agent is built, publish it to make the changes live, and run a conversation against it to see how it behaves.