Files
adk-python/contributing/samples/workflow_triage
Wei Sun (Jack) e295feb4c6 docs: Add workflow_triage sample for multi-agent request orchestration
Demonstrates intelligent triage system with root planning agent and parallel
execution agents.

Use session state to store the execution plan and coordinate with other specialized agents.

Check out README.md for more details.

PiperOrigin-RevId: 793884758
2025-08-11 19:05:53 -07:00
..

Workflow Triage Sample

This sample demonstrates how to build a multi-agent workflow that intelligently triages incoming requests and delegates them to appropriate specialized agents.

Overview

The workflow consists of three main components:

  1. Execution Manager Agent (agent.py) - Analyzes user input and determines which execution agents are relevant
  2. Plan Execution Agent - Sequential agent that coordinates execution and summarization
  3. Worker Execution Agents (execution_agent.py) - Specialized agents that execute specific tasks in parallel

Architecture

Execution Manager Agent (root_agent)

  • Model: gemini-2.5-flash
  • Name: execution_manager_agent
  • Role: Analyzes user requests and updates the execution plan
  • Tools: update_execution_plan - Updates which execution agents should be activated
  • Sub-agents: Delegates to plan_execution_agent for actual task execution
  • Clarification: Asks for clarification if user intent is unclear before proceeding

Plan Execution Agent

  • Type: SequentialAgent
  • Name: plan_execution_agent
  • Components:
    • worker_parallel_agent (ParallelAgent) - Runs relevant agents in parallel
    • execution_summary_agent - Summarizes the execution results

Worker Agents

The system includes two specialized execution agents that run in parallel:

  • Code Agent (code_agent): Handles code generation tasks
    • Uses before_agent_callback_check_relevance to skip if not relevant
    • Output stored in code_agent_output state key
  • Math Agent (math_agent): Performs mathematical calculations
    • Uses before_agent_callback_check_relevance to skip if not relevant
    • Output stored in math_agent_output state key

Execution Summary Agent

  • Model: gemini-2.5-flash
  • Name: execution_summary_agent
  • Role: Summarizes outputs from all activated agents
  • Dynamic Instructions: Generated based on which agents were activated
  • Content Inclusion: Set to "none" to focus on summarization

Key Features

  • Dynamic Agent Selection: Automatically determines which agents are needed based on user input
  • Parallel Execution: Multiple relevant agents can work simultaneously via ParallelAgent
  • Relevance Filtering: Agents skip execution if they're not relevant to the current state using callback mechanism
  • Stateful Workflow: Maintains execution state through ToolContext
  • Execution Summarization: Automatically summarizes results from all activated agents
  • Sequential Coordination: Uses SequentialAgent to ensure proper execution flow

Usage

The workflow follows this pattern:

  1. User provides input to the root agent (execution_manager_agent)
  2. Manager analyzes the request and identifies relevant agents (code_agent, math_agent)
  3. If user intent is unclear, manager asks for clarification before proceeding
  4. Manager updates the execution plan using update_execution_plan
  5. Control transfers to plan_execution_agent
  6. worker_parallel_agent (ParallelAgent) runs only relevant agents based on the updated plan
  7. execution_summary_agent summarizes the results from all activated agents

Example Queries

Vague requests requiring clarification:

> hi
> Help me do this.

The root agent (execution_manager_agent) will greet the user and ask for clarification about their specific task.

Math-only requests:

> What's 1+1?

Only the math_agent executes while code_agent is skipped.

Multi-domain requests:

> What's 1+11? Write a python function to verify it.

Both code_agent and math_agent execute in parallel, followed by summarization.

Available Execution Agents

  • code_agent - For code generation and programming tasks
  • math_agent - For mathematical computations and analysis

Implementation Details

  • Uses Google ADK agents framework
  • Implements callback-based relevance checking via before_agent_callback_check_relevance
  • Maintains state through ToolContext and state keys
  • Supports parallel agent execution with ParallelAgent
  • Uses SequentialAgent for coordinated execution flow
  • Dynamic instruction generation for summary agent based on activated agents
  • Agent outputs stored in state with {agent_name}_output keys