Sequential vs Hierarchical Crew Processes

CrewAI's process field chooses how a crew schedules work: a fixed pipeline (Process.sequential) or a manager-led team (Process.hierarchical).

Use this cheatsheet when you decide between static ownership and dynamic delegation.

How to Use This Cheatsheet

Start from section A if you are choosing a default for a new crew.
Use B when implementing hierarchical managers and validation loops.
Use C when debugging cost, thrash, or unclear ownership.
Prefer sequential until a product requirement forces dynamic assignment or manager review.

A - Decision Snapshot

Criterion	Sequential	Hierarchical
Default	Yes (`Process.sequential`)	Opt-in (`Process.hierarchical`)
Task order	Fixed list order	Manager plans / assigns
Task owner	Set via `task.agent`	Manager allocates by capability
Extra config	None	Requires `manager_llm` or `manager_agent`
Best for	Known pipelines (research → write)	Variable routing, review, rework
Cost profile	One pass per task (plus tools)	Manager tokens + possible rework
Debuggability	High (linear logs)	Medium (delegation noise)
Failure mode	Weak handoff / vague expected_output	Manager thrash / mis-assignment

Rule of thumb: if you can draw the pipeline on one line, use sequential. If a human PM would reassign tickets mid-project, consider hierarchical.

B - Implementation Cheatsheet

Sequential essentials

from crewai import Crew, Process
 
crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, write_task],  # order matters
    process=Process.sequential,
)

Knob	Guidance
Task list order	Defines execution order
`Task.context`	Explicit upstream outputs (do not rely on vibes)
`Task.agent`	Required ownership clarity
`planning=True`	Optional pre-pass; still sequential execution

Hierarchical essentials

from crewai import Crew, Process
 
crew = Crew(
    agents=[researcher, writer, reviewer],
    tasks=[research_task, write_task, review_task],
    process=Process.hierarchical,
    manager_llm="gpt-4o",  # or manager_agent=custom_manager
)

Knob	Guidance
`manager_llm`	Auto manager powered by this model
`manager_agent`	Custom manager with its own role/goal/tools
Worker roles	Must be distinguishable or assignment is random theater
Validation	Manager reviews outcomes before proceeding - budget for it

Side-by-side kickoff

Step	Sequential	Hierarchical
1	Run task 1 with its agent	Manager interprets backlog
2	Feed outputs forward	Delegate to workers
3	Run task N	Review / possibly rework
4	Return `CrewOutput`	Return `CrewOutput`

C - Choose Sequential When

The workflow is a stable assembly line (ingest → analyze → draft → format).
Each task has an obvious owner and tool set.
You need cheap, linear logs for evals and CI.
Latency and token budgets are tight.
You can express dependencies with context=[...].

D - Choose Hierarchical When

Work must be routed based on intermediate findings.
You need a reviewer/manager role that is not just another pipeline step.
Task ownership should change without rewriting the task list every time.
Specialists are many and the "who does this?" decision is non-trivial.
You accept higher cost for coordination quality.

E - Hybrid Patterns (Without Lying to Yourself)

Pattern	How	Caution
Sequential + planning	`planning=True` on sequential crew	Planning annotates; it is not full management
Sequential review task	Final critic agent with fixed owner	Cheaper than hierarchical; less flexible
Hierarchical for phase 1 only	Separate crews	Clearer than one mega-crew with mixed modes
Flow orchestration	CrewAI Flows around crews	Use when product needs state/chat, not only process enum

F - Cost, Risk, and Ops

Concern	Sequential	Hierarchical
Token cost	Lower baseline	Manager + possible retries
Latency	Predictable stages	Variable
Eval design	Per-task golden outputs	Include manager decision quality
Stuck loops	Bad task specs / tools	Bad manager prompts + worker overlap
Observability	Task callbacks enough	Log delegations and validation verdicts

G - Quick Failure Diagnosis

Symptom	Likely process issue	Fix
Writer ignores research	Missing `context` or weak expected_output	Sequential handoff repair
Endless "I'll reassign" chatter	Hierarchical thrash	Sharper worker goals; better manager goal; caps
Same agent does everything	Overlapping roles	Differentiate tools and goals
Bill shock	Manager rework loops	Cap iterations; switch hot paths to sequential
Non-deterministic owners	Hierarchical without distinct skills	Make roles non-substitutable

FAQs

Is sequential always faster?

Usually yes for the same task set, because you skip manager planning and review turns. Tool latency can still dominate either mode.

Can I set process per task?

Process is a crew-level strategy. Split into multiple crews or use Flows if phases need different coordination styles.

Do hierarchical crews ignore task.agent?

Treat hierarchical as manager-driven allocation and validation. Design workers so the manager has clear choices; verify current assignment behavior at build for your version.

Is manager_agent better than manager_llm?

Use manager_agent when you need a crafted manager role, tools, or model. Use manager_llm for a quick default manager.

Does allow_delegation replace hierarchical?

No. Delegation is a worker capability. Hierarchical is a crew process with an explicit manager.

Should demos use hierarchical to look impressive?

No. Demos should use sequential unless you are showcasing management itself.

How does planning interact with hierarchical?

Planning can add a pre-pass; hierarchical still owns coordination. Measure combined cost before enabling both.

What process should CI use?

Sequential with fixed seeds/inputs when possible. Hierarchical evals need extra assertions on assignment and review.

Can hierarchical fix vague tasks?

No. Managers amplify ambiguity. Clarify description and expected_output first.

When should I leave Crew processes for Flows?

When you need multi-turn chat sessions, branching product logic, or durable state beyond a single crew kickoff.

The CrewAI Mental Model: Roles, Tasks, and Crews - building blocks
CrewAI Basics - first sequential crew
Defining Agent Roles, Goals, and Backstories - worker clarity
Debugging a Crew That Won't Converge on an Answer - thrash fixes
CrewAI Best Practices - scoping checklist
The CrewAI Chat API for Interactive Crew Sessions - multi-turn product surface

Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana @solana/web3.js + @solana/spl-token.

Sequential vs Hierarchical Crew Processes

CrewAI's process field chooses how a crew schedules work: a fixed pipeline (Process.sequential) or a manager-led team (Process.hierarchical).

Use this cheatsheet when you decide between static ownership and dynamic delegation.

How to Use This Cheatsheet

Start from section A if you are choosing a default for a new crew.
Use B when implementing hierarchical managers and validation loops.
Use C when debugging cost, thrash, or unclear ownership.
Prefer sequential until a product requirement forces dynamic assignment or manager review.

A - Decision Snapshot

Criterion	Sequential	Hierarchical
Default	Yes (`Process.sequential`)	Opt-in (`Process.hierarchical`)
Task order	Fixed list order	Manager plans / assigns
Task owner	Set via `task.agent`	Manager allocates by capability
Extra config	None	Requires `manager_llm` or `manager_agent`
Best for	Known pipelines (research → write)	Variable routing, review, rework
Cost profile	One pass per task (plus tools)	Manager tokens + possible rework
Debuggability	High (linear logs)	Medium (delegation noise)
Failure mode	Weak handoff / vague expected_output	Manager thrash / mis-assignment

Rule of thumb: if you can draw the pipeline on one line, use sequential. If a human PM would reassign tickets mid-project, consider hierarchical.

B - Implementation Cheatsheet

Sequential essentials

from crewai import Crew, Process
 
crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, write_task],  # order matters
    process=Process.sequential,
)

Knob	Guidance
Task list order	Defines execution order
`Task.context`	Explicit upstream outputs (do not rely on vibes)
`Task.agent`	Required ownership clarity
`planning=True`	Optional pre-pass; still sequential execution

Hierarchical essentials

from crewai import Crew, Process
 
crew = Crew(
    agents=[researcher, writer, reviewer],
    tasks=[research_task, write_task, review_task],
    process=Process.hierarchical,
    manager_llm="gpt-4o",  # or manager_agent=custom_manager
)

Knob	Guidance
`manager_llm`	Auto manager powered by this model
`manager_agent`	Custom manager with its own role/goal/tools
Worker roles	Must be distinguishable or assignment is random theater
Validation	Manager reviews outcomes before proceeding - budget for it

Side-by-side kickoff

Step	Sequential	Hierarchical
1	Run task 1 with its agent	Manager interprets backlog
2	Feed outputs forward	Delegate to workers
3	Run task N	Review / possibly rework
4	Return `CrewOutput`	Return `CrewOutput`

C - Choose Sequential When

The workflow is a stable assembly line (ingest → analyze → draft → format).
Each task has an obvious owner and tool set.
You need cheap, linear logs for evals and CI.
Latency and token budgets are tight.
You can express dependencies with context=[...].

D - Choose Hierarchical When

Work must be routed based on intermediate findings.
You need a reviewer/manager role that is not just another pipeline step.
Task ownership should change without rewriting the task list every time.
Specialists are many and the "who does this?" decision is non-trivial.
You accept higher cost for coordination quality.

E - Hybrid Patterns (Without Lying to Yourself)

Pattern	How	Caution
Sequential + planning	`planning=True` on sequential crew	Planning annotates; it is not full management
Sequential review task	Final critic agent with fixed owner	Cheaper than hierarchical; less flexible
Hierarchical for phase 1 only	Separate crews	Clearer than one mega-crew with mixed modes
Flow orchestration	CrewAI Flows around crews	Use when product needs state/chat, not only process enum

F - Cost, Risk, and Ops

Concern	Sequential	Hierarchical
Token cost	Lower baseline	Manager + possible retries
Latency	Predictable stages	Variable
Eval design	Per-task golden outputs	Include manager decision quality
Stuck loops	Bad task specs / tools	Bad manager prompts + worker overlap
Observability	Task callbacks enough	Log delegations and validation verdicts

G - Quick Failure Diagnosis

Symptom	Likely process issue	Fix
Writer ignores research	Missing `context` or weak expected_output	Sequential handoff repair
Endless "I'll reassign" chatter	Hierarchical thrash	Sharper worker goals; better manager goal; caps
Same agent does everything	Overlapping roles	Differentiate tools and goals
Bill shock	Manager rework loops	Cap iterations; switch hot paths to sequential
Non-deterministic owners	Hierarchical without distinct skills	Make roles non-substitutable

FAQs

Is sequential always faster?

Usually yes for the same task set, because you skip manager planning and review turns. Tool latency can still dominate either mode.

Can I set process per task?

Process is a crew-level strategy. Split into multiple crews or use Flows if phases need different coordination styles.

Do hierarchical crews ignore task.agent?

Treat hierarchical as manager-driven allocation and validation. Design workers so the manager has clear choices; verify current assignment behavior at build for your version.

Is manager_agent better than manager_llm?

Use manager_agent when you need a crafted manager role, tools, or model. Use manager_llm for a quick default manager.

Does allow_delegation replace hierarchical?

No. Delegation is a worker capability. Hierarchical is a crew process with an explicit manager.

Should demos use hierarchical to look impressive?

No. Demos should use sequential unless you are showcasing management itself.

How does planning interact with hierarchical?

Planning can add a pre-pass; hierarchical still owns coordination. Measure combined cost before enabling both.

What process should CI use?

Sequential with fixed seeds/inputs when possible. Hierarchical evals need extra assertions on assignment and review.

Can hierarchical fix vague tasks?

No. Managers amplify ambiguity. Clarify description and expected_output first.

When should I leave Crew processes for Flows?

When you need multi-turn chat sessions, branching product logic, or durable state beyond a single crew kickoff.

The CrewAI Mental Model: Roles, Tasks, and Crews - building blocks
CrewAI Basics - first sequential crew
Defining Agent Roles, Goals, and Backstories - worker clarity
Debugging a Crew That Won't Converge on an Answer - thrash fixes
CrewAI Best Practices - scoping checklist
The CrewAI Chat API for Interactive Crew Sessions - multi-turn product surface

Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana @solana/web3.js + @solana/spl-token.

Sequential vs Hierarchical Crew Processes

Buscar en la documentación

Sequential vs Hierarchical Crew Processes