Hierarchical Process

In CrewAI, a hierarchical process adds a manager layer on top of your specialist agents. The manager coordinates the run, delegates work to the right people based on what each task needs, and checks that outputs meet expectations before the crew finishes. Specialists focus on execution; the manager focuses on planning, assignment, and quality control.

Tasks are not pre-assigned to agents

Unlike a rigid sequential pipeline where you often pin each Task to a specific agent, hierarchical mode is built around delegation: you still define what needs to be done (task descriptions and expected outputs), but who does each piece is not fixed up front in the same way. The manager reads the task, considers each specialist’s role, goal, and capabilities, and allocates work at runtime. That is why hierarchical crews shine when the best next step depends on intermediate results rather than a fixed script.

Two ways to configure the manager

1. Auto-created manager — manager_llm

Pass an LLM string (or compatible config). CrewAI creates a manager for you that uses that model to reason about delegation and validation.

crew = Crew(
    agents=[researcher, analyst, writer],
    tasks=[task_a, task_b, task_c],
    process=Process.hierarchical,
    manager_llm="gpt-4o",
    verbose=True,
)

2. Custom manager — manager_agent

Build an Agent with an explicit role, goal, and backstory so the lead behaves like your product owner, tech lead, or editor—not a generic coordinator.

crew = Crew(
    agents=[researcher, analyst, writer],
    tasks=[task_a, task_b, task_c],
    process=Process.hierarchical,
    manager_agent=project_manager,
    verbose=True,
)

Use manager_llm for a quick start; use manager_agent when tone, priorities, and decision style matter for your product.

Complete example: three specialists and Process.hierarchical

Below, three specialists (research, analysis, writing) are available to the crew. Tasks describe outcomes; under hierarchical process the manager decides how to use the team. Worker agents typically set allow_delegation=False so only the manager orchestrates. The manager (auto or custom) uses allow_delegation=True when you configure it explicitly.

from crewai import Agent, Task, Crew, Process
 
researcher = Agent(
    role="Researcher",
    goal="Gather accurate, sourced facts and concise notes on the topic.",
    backstory="You are rigorous about facts and avoid speculation.",
    verbose=True,
    allow_delegation=False,
)
 
analyst = Agent(
    role="Analyst",
    goal="Turn raw information into clear insights, risks, and recommendations.",
    backstory="You structure messy inputs and call out assumptions.",
    verbose=True,
    allow_delegation=False,
)
 
writer = Agent(
    role="Writer",
    goal="Produce polished prose that matches the requested format.",
    backstory="You edit for clarity and audience-appropriate tone.",
    verbose=True,
    allow_delegation=False,
)
 
task_a = Task(
    description="Understand the landscape for {topic}: key players, trends, and one open question.",
    expected_output="Short structured notes suitable for further analysis.",
)
 
task_b = Task(
    description="From the team’s findings so far, identify top risks and opportunities for {topic}.",
    expected_output="Bulleted risks and opportunities with one-line rationale each.",
)
 
task_c = Task(
    description="Produce a brief executive summary for stakeholders on {topic}.",
    expected_output="Two or three tight paragraphs, no fluff.",
)
 
crew = Crew(
    agents=[researcher, analyst, writer],
    tasks=[task_a, task_b, task_c],
    process=Process.hierarchical,
    manager_llm="gpt-4o",
    verbose=True,
)
 
result = crew.kickoff(inputs={"topic": "agentic AI in enterprises"})
print(result.raw)

Manager responsibilities

• Task delegation — Breaks work down (implicitly or explicitly), picks which specialist handles each slice, and sequences or parallelizes based on the situation.
• Result validation — Compares worker outputs to expected_output and the overall goal; may send work back for revision or ask another agent to supplement.
• Conflict resolution — When two perspectives disagree or quality is unclear, the manager arbitrates, asks for clarification, or reframes the next task so the crew converges.

Hierarchical vs sequential

Sequential crews are usually easier to debug line-by-line; hierarchical crews trade that for flexibility when the manager’s judgment adds real value.

Custom manager agent: role and goal

Define the manager like any other agent, but aim its role and goal at coordination and quality, not at doing every specialist job itself. Keep worker agents in the agents list; pass the manager only via manager_agent (do not duplicate it as a worker unless your version’s docs explicitly require otherwise).

project_manager = Agent(
    role="Engagement Lead",
    goal="Deliver a coherent, client-ready brief by delegating to research, analysis, and writing experts—and reject work that misses the brief.",
    backstory=(
        "You have run consulting engagements for a decade. "
        "You insist on clarity, traceable reasoning, and one voice in the final summary."
    ),
    verbose=True,
    allow_delegation=True,
)
 
crew = Crew(
    agents=[researcher, analyst, writer],
    tasks=[task_a, task_b, task_c],
    process=Process.hierarchical,
    manager_agent=project_manager,
    verbose=True,
)

Key takeaways

• Hierarchical process uses a manager to coordinate, delegate to specialists, and validate outcomes.
• Delegation is dynamic: tasks describe work; assignment follows capabilities and runtime context, not a single fixed mapping you set in stone.
• Configure the lead with manager_llm="gpt-4o" (auto manager) or manager_agent=Agent(...) (custom personality and priorities).
• Use Process.hierarchical on Crew together with one of those manager options.
• Choose hierarchical for complex, context-dependent workflows; choose sequential for straightforward, ordered pipelines.