Skip to main content

​
Description

Agent Composer is a framework that allows you to compose custom query path implementations using pre-made building blocks. By combining primitives such as search tools, third-party API/MCP tools, and various LLMs, you can construct arbitrary computational graphs and configure them to build specialized agents.
Template agents (Basic Search and Agentic Search) are available for self-serve users. Full Agent Composer capabilities—including the visual workflow builder for custom workflows, YAML customization, and the prompt-based workflow generator—are available in public preview for enterprise users.

Read the launch announcement

Learn how Agent Composer reduces complex technical tasks from hours to minutes.

​
Two Ways to Build

The framework offers two ways to construct workflows:
  1. YAML Configuration: Specify a YAML with a specific format describing the graph and its configuration. Enter this YAML in the “Agent Composer” tab during agent creation.
  2. Visual Workflow Builder: Use the visual frontend builder, which internally translates to YAML.

​
Key Concepts

ConceptDescription
NodesBasic operations in the graph (search, generate, transform, etc.)
Input MappingHow nodes connect to each other using source_node#output_key syntax
ConfigConfiguration parameters passed to step constructors
SubgraphsReusable components that can be referenced in larger workflows
Conditional NodesExecute different paths based on runtime conditions
UI Stream TypesControl what data streams to the UI during execution

​
Example: Minimal Agent

version: 0.1
inputs:
  query: str

outputs:
  response: str

nodes:
  search:
    type: SearchUnstructuredDataStep
    input_mapping:
      query: __inputs__#query

  generate:
    type: ResponseGenerationStep
    input_mapping:
      query: __inputs__#query
      retrievals: search#retrievals

  __outputs__:
    type: output
    input_mapping:
      response: generate#response
This minimal example:
  • Accepts a query string as input
  • Searches the datastore for relevant documents
  • Generates a response based on the query and retrieved documents
  • Returns the response as output