Implementation:Microsoft Agent framework DeepResearch Workflow Definition

Knowledge Sources	Microsoft_Agent_framework Magentic-One
Domains	Multi_Agent_Orchestration, Declarative_Workflows, Deep_Research
Last Updated	2026-02-11 17:00 GMT

Overview

Sample declarative YAML workflow implementing the Magentic orchestration pattern to coordinate multiple specialized agents for deep research tasks.

Description

The DeepResearch.yaml workflow defines a complex multi-agent orchestration that addresses user requests through coordinated agent collaboration. It uses seven specialized agents organized into two tiers:

Coordination agents:

ResearchAgent: Analyzes the current task and correlates relevant facts
PlannerAgent: Devises an overall plan based on facts and team capabilities
ManagerAgent: Evaluates progress via a "progress ledger" and delegates tasks
SummaryAgent: Generates the final response when the request is satisfied

Capability agents:

KnowledgeAgent: Performs generic web searches
CoderAgent: Writes and executes Python code
WeatherAgent: Provides weather information

The workflow implements a loop with stall detection: if the ManagerAgent determines progress has stalled or the system is in a loop, it re-analyzes facts and re-plans. After exceeding a stall threshold, it restarts entirely. After exceeding a restart threshold, it terminates gracefully.

Usage

Use this workflow definition as a reference pattern when building complex multi-agent orchestrations that require iterative planning, fact-gathering, and dynamic task delegation with failure recovery.

Code Reference

Source Location

Repository: Microsoft_Agent_framework
File: workflow-samples/DeepResearch.yaml
Lines: 1-379

Signature

kind: Workflow
trigger:
  kind: OnConversationStart
  id: workflow_demo
  actions:
    # 1. Initialize available agents list
    - kind: SetVariable
      variable: Local.AvailableAgents
      value: =[{name: "WeatherAgent", ...}, {name: "CoderAgent", ...}, {name: "KnowledgeAgent", ...}]

    # 2. Gather facts via ResearchAgent
    - kind: InvokeAzureAgent
      agent:
        name: ResearchAgent

    # 3. Create plan via PlannerAgent
    - kind: InvokeAzureAgent
      agent:
        name: PlannerAgent

    # 4. Main loop: ManagerAgent evaluates progress ledger
    - kind: InvokeAzureAgent
      agent:
        name: ManagerAgent
      output:
        responseObject: Local.ProgressLedger

    # 5. Conditional branching: done / stalling / delegate
    - kind: ConditionGroup
      conditions:
        - condition: =Local.ProgressLedger.is_request_satisfied.answer
        - condition: =Local.ProgressLedger.is_in_loop.answer || Not(Local.ProgressLedger.is_progress_being_made.answer)

    # 6. Delegate to next agent and loop back
    - kind: GotoAction
      actionId: question_o3BQkf

Import

from agent_framework_declarative import AgentFactory

# Load the DeepResearch workflow from YAML
factory = AgentFactory(...)
agent = await factory.create_agent_from_yaml("workflow-samples/DeepResearch.yaml")

I/O Contract

Inputs

Name	Type	Required	Description
System.LastMessage.Text	string	Yes	The user's research query that triggers the workflow
Agent definitions	Azure agents	Yes	ResearchAgent, PlannerAgent, ManagerAgent, SummaryAgent, WeatherAgent, CoderAgent, KnowledgeAgent must be configured

Outputs

Name	Type	Description
Final response	ChatMessage	Generated by SummaryAgent after ManagerAgent determines request is satisfied
Progress activities	SendActivity	Intermediate status updates sent during research (fact analysis, plan creation, agent delegation)

Usage Examples

Running the DeepResearch Workflow

from agent_framework_declarative import AgentFactory
from agent_framework_core import Agent

# Set up the factory with required agents
factory = AgentFactory(
    client=azure_client,
    tools=[],
)

# Load and run the workflow
agent = await factory.create_agent_from_yaml("workflow-samples/DeepResearch.yaml")
async with agent:
    response = await agent.run("What are the latest advances in quantum computing?")
    async for event in response:
        print(event)

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment