Description

Description

In CrewAI's hierarchical process, manager agents are unable to delegate tasks to worker agents, even when allow_delegation=True is set. This breaks the core functionality of hierarchical task coordination.

Expected Behavior

In hierarchical process:

1. Manager agent should analyze tasks and delegate subtasks to appropriate worker agents
2. Worker agents should execute delegated tasks
3. Manager should coordinate and combine results

Actual Behavior

Manager agents only execute tasks using their own tools and never delegate to other agents, effectively making the hierarchical process behave like a sequential process.

Steps to Reproduce

1. Create a Crew with hierarchical process: from crewai import Crew, Process, Agent, Task

Create worker agents

worker1 = Agent( role="Data Analyst", goal="Analyze data and provide insights", tools=[some_data_analysis_tool], allow_delegation=False )

worker2 = Agent( role="Report Writer", goal="Write reports based on analysis", tools=[some_writing_tool], allow_delegation=False )

Create manager agent (no pre-defined manager agent)

manager_llm = some_llm_instance

Create crew

crew = Crew( agents=[worker1, worker2], tasks=[complex_analysis_task], process=Process.hierarchical, # This should enable delegation manager_llm=manager_llm, verbose=True )

Execute

result = crew.kickoff()

Steps to Reproduce

Steps to Reproduce

Step 1: Set up a simple test environment

Create a new Python file test_delegation_bug.py:

#!/usr/bin/env python3 """ Test script to reproduce CrewAI hierarchical delegation bug """

from crewai import Crew, Process, Agent, Task, LLM

def test_delegation_bug(): """Reproduce the delegation bug in hierarchical process"""

# Create a simple mock LLM for testing (replace with your actual LLM)
llm = LLM(model="gpt-4", temperature=0.1)  # Adjust based on your setup

# Step 2: Create worker agents with specialized tools
researcher = Agent(
    role="Research Specialist",
    goal="Conduct research and gather information",
    backstory="You are an expert researcher who can find and analyze information.",
    tools=[],  # No special tools for this test
    allow_delegation=False,  # Worker agents shouldn't delegate
    verbose=True,
    llm=llm
)

writer = Agent(
    role="Content Writer",
    goal="Write clear and engaging content based on research",
    backstory="You are a skilled writer who creates compelling content.",
    tools=[],  # No special tools for this test
    allow_delegation=False,  # Worker agents shouldn't delegate
    verbose=True,
    llm=llm
)

# Step 3: Create a complex task that should be delegated
research_task = Task(
    description="""
    Create a comprehensive article about artificial intelligence trends in 2024.
    
    This task requires:
    1. Research current AI trends and technologies
    2. Analyze market impacts and future predictions  
    3. Write a well-structured article
    
    Delegate the research work to the Research Specialist and writing to the Content Writer.
    """,
    expected_output="A complete article about AI trends in 2024",
    agent=None  # No specific agent assigned - should go to manager
)

# Step 4: Create crew with hierarchical process (THIS IS WHERE THE BUG OCCURS)
crew = Crew(
    agents=[researcher, writer],  # Only worker agents
    tasks=[research_task],
    process=Process.hierarchical,  # Manager should coordinate
    manager_llm=llm,  # Manager LLM provided, no pre-defined manager_agent
    verbose=True
)

print("=== CrewAI Hierarchical Delegation Test ===")
print(f"Process: {crew.process}")
print(f"Number of agents: {len(crew.agents)}")
print(f"Manager agent: {crew.manager_agent}")  # Should be None initially
print()

# Step 5: Execute and observe the bug
print("Executing crew.kickoff()...")
print("Expected: Manager delegates research to researcher, writing to writer")
print("Actual bug: Manager does all work itself, no delegation occurs")
print("-" * 60)

try:
    result = crew.kickoff()
    print(f"\nResult: {result}")
    
    # Check if delegation actually occurred
    print(f"\nDelegation tracking:")
    print(f"Manager agent after execution: {crew.manager_agent}")
    if hasattr(crew.manager_agent, 'tools'):
        print(f"Manager tools count: {len(crew.manager_agent.tools or [])}")
    
except Exception as e:
    print(f"Error occurred: {e}")
    import traceback
    traceback.print_exc()

if name == "main": test_delegation_bug()

Expected behavior

Expected behavior

When using CrewAI's hierarchical process with dynamically created manager agents, the manager should be able to properly delegate tasks to worker agents instead of performing all work itself.

Specific Expected Behaviors:

1. Successful Delegation:

   • Manager agent calls "Delegate work to coworker" tool
   • Tool successfully finds and delegates to worker agents
   • No "No agent found" errors for valid coworkers

2. Parallel Task Execution:

• Manager Agent → Delegates to Worker A → Worker A executes
• → Delegates to Worker B → Worker B executes
• → Coordinates results

1. Proper Tool Configuration:

   • Delegation tools should target other agents in the crew
   • Manager should not attempt to delegate to itself
   • Worker agents should receive delegated tasks

2. Observable Results:

   • task.delegations counter should be > 0
   • Worker agents should show execution logs
   • Manager should coordinate rather than execute all subtasks
   • Tasks should complete through proper delegation chain

Example Expected Console Output:

[Manager Agent] Analyzing: "Create article about AI trends" [Manager Agent] Delegating research task to Research Specialist... [Research Specialist] Executing research on AI trends... [Manager Agent] Delegating writing task to Content Writer... [Content Writer] Writing article based on research... [Manager Agent] Coordinating final article... ✅ Task completed with proper delegation

Screenshots/Code snippets

Fixed Code (proposed): def _update_manager_tools( self, task: Task, tools: list[BaseTool] ) -> list[BaseTool]: if self.manager_agent: if task.agent: # FIX: Allow delegation to other agents (excluding self) other_agents = [agent for agent in self.agents if agent != task.agent] tools = self._inject_delegation_tools(tools, task.agent, other_agents) else: tools = self._inject_delegation_tools( tools, self.manager_agent, self.agents ) return tools

=== CrewAI Hierarchical Delegation Test === Process: Process.hierarchical Number of agents: 2 Manager agent: None

Executing crew.kickoff()... Expected: Manager delegates research to researcher, writing to writer Actual bug: Manager does all work itself, no delegation occurs

[Manager Agent] Starting task execution... [Manager Agent] Using tool: some_analysis_tool (manager doing research itself) [Manager Agent] Using tool: some_writing_tool (manager doing writing itself) [Manager Agent] Task completed - no delegation occurred!

Delegation tracking: Manager agent after execution: <Agent role='Manager' ...> Manager tools count: 2 # Has delegation tools but they don't work

Operating System

macOS Ventura

Python Version

3.10

crewAI Version

1.10.1

crewAI Tools Version

CrewAI Tools: Not installed / Not used

Virtual Environment

Venv

Evidence

Evidence

1. Debug Logs Showing the Bug

Console output from test script (run the reproduction script):

Possible Solution

https://github.com/crewAIInc/crewAI/pull/4782

Additional context

None