Hierarchical wins when task routing must depend on intermediate results an LLM evaluates. Sequential is the right default whenever the order is known at design time.
Picture a kitchen. Sequential is a fixed recipe. Chop, sauté, plate, in that order, every time. Hierarchical hires a head chef who watches what is going on, decides who chops next, sends a sous chef back to re-sear something, and only declares the dish done when the plate looks right. The fixed recipe is faster and cheaper when the dish never changes. The head chef earns their salary when the order of moves depends on what is happening on the stove. Most weeknight cooking is recipes. Catering complicated menus is when you want a chef.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
CrewAI's Process parameter is a small API surface with a large architectural consequence. Sequential gives you a declarative DAG. Hierarchical gives you runtime dispatch through a manager LLM. Picking the wrong one is rarely a correctness bug, both will eventually return something, but it shows up as cost, latency, and reliability differences that compound under production load.
This section walks the mechanical difference, names the workloads where each earns its keep, lays out the cost model honestly, and connects the Hierarchical choice to its natural successor in the framework landscape: LangGraph's supervisor pattern.
What Sequential actually does
Sequential is the simpler primitive. You declare a list of Task objects, each with an agent assigned, in the order they should run. The Crew iterates through the list. Each task's output is appended to a shared context string that gets passed to the next task. The flow is fixed at construction time: changing the order means editing the code.
This sounds limiting but is actually the right choice for most production workloads. ETL-shaped flows (fetch, summarize, classify, write) have a stable DAG. Report generation pipelines have a stable DAG. Code-review workflows where you always lint, then type-check, then run tests have a stable DAG. For any flow you would draw the same way every time on a whiteboard, Sequential is the correct primitive and the determinism is a feature, not a limitation.
The operational properties follow from this: Sequential flows are easy to test (the order is known), easy to trace (each step has a fixed identity), and easy to cost-model (N tasks means N LLM calls). The downside is rigidity. Any branching, any 'this depends on what the last step found', anything dynamic. Sequential cannot express it without you writing the routing logic outside the framework.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
| Aspect | Sequential | Hierarchical |
|---|---|---|
| Routing decision | Author writes it at design time | Manager LLM decides at run time |
| Cost per step | Worker LLM call only | Worker LLM call plus manager dispatch call |
| Determinism | High. Same DAG every run | Low. Manager output varies |
| Best fit | Known pipelines, ETL, report generation | Exploratory research, iterative refinement, dynamic delegation |
| Failure mode | Wrong task order in code | Manager loops or routes to wrong agent |
Real products, models, and research that use this idea.
- CrewAI's official examples use Sequential for research to report flows where the DAG is fixed
- Hierarchical shows up in CrewAI's exploratory-research templates where downstream tasks depend on what the researcher found
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does Hierarchical Process handle termination, and what happens if you forget to bound it?
The manager keeps dispatching until it concludes the goal is met. Without a max_iter or explicit termination signal in the goal description, you can get a runaway loop. Production teams set hard caps and monitor manager output for loop signatures.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Reaching for Hierarchical because it 'sounds smarter' when the task DAG is fully known at design time. You are paying for an extra LLM call per routing decision with no benefit.
60 second bullets to scan on the way to the call.
Definition of Sequential Process and its DAG semantics
Definition of Hierarchical Process and the role of manager_llm or manager_agent
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.