Sequential Pipeline

A multi-agent pipeline that chains two sub-agents: a researcher and a writer.

schema_version: "1.0.0"

metadata:
  id: "research-pipeline"
  name: "Research Pipeline"
  version: "1.0.0"
  description: "A sequential multi-agent pipeline"

interface:
  input:
    type: object
    properties:
      query:
        type: string
    required: [query]
  output:
    type: object
    properties:
      result:
        type: string
    required: [result]

execution_policy:
  id: agf.sequential
  config:
    steps:
      - agent: researcher
        input_mapping:
          query: "parent.input.query"
      - agent: writer
        input_mapping:
          research: "researcher.output.result"
          query: "parent.input.query"
    output_from: "last"

action_space:
  local_agents:
    - alias: researcher
      source: "./agents/researcher.agf.yaml"
      description: "Researches the topic"
    - alias: writer
      source: "./agents/writer.agf.yaml"
      description: "Writes the final response"

Key Concepts

Sequential Policy

The agf.sequential policy runs sub-agents in order. Each step completes before the next begins.

execution_policy:
  id: agf.sequential
  config:
    steps:
      - agent: researcher
        input_mapping:
          query: "parent.input.query"
      - agent: writer
        input_mapping:
          research: "researcher.output.result"
          query: "parent.input.query"
    output_from: "last"

Field	Description
steps[].agent	Alias of the sub-agent (must match a local_agents entry)
steps[].input_mapping	Maps data from parent input or previous agent outputs
output_from	Which step's output becomes the pipeline output ("last" or a step alias)

Input Mapping Expressions

Input mappings use dot-notation paths:

Expression	Meaning
parent.input.query	The pipeline's own input field query
researcher.output.result	The result field from the researcher's output

This lets you wire data between agents without code.

Local Agents

action_space:
  local_agents:
    - alias: researcher
      source: "./agents/researcher.agf.yaml"
      description: "Researches the topic"

Each sub-agent is defined in its own .agf.yaml file and referenced by relative path. The agf graph command visualizes these relationships:

info

The agf CLI will be publicly available soon. See the CLI Reference for the command reference. You can validate this example now using the Playground.

agf graph research-pipeline.agf.yaml
# research-pipeline
# ├── researcher (./agents/researcher.agf.yaml)
# └── writer (./agents/writer.agf.yaml)

Other Orchestration Policies

Agent Format supports several orchestration patterns beyond sequential:

Policy	Description
agf.sequential	Run steps in order
agf.parallel	Run branches concurrently, merge results
agf.loop	Repeat until a condition is met
agf.batch	Apply the same agent to each item in a list
agf.conditional	Route to different agents based on conditions