When does CrewAI hierarchical Process beat sequential?

CrewAI lets you wire a set of Agents to a list of Tasks. The `Process` parameter chooses how those Tasks get executed. **Sequential** is a pipeline. The Tasks run in the order you declared them. Each Task is assigned to a specific Agent. The output of one Task is threaded into the context of the next. There is no decision-making at run time. You decided the order when you wrote the code. **Hierarchical** adds a manager. You supply either a `manager_llm` (a model that plays the manager role) or a full `manager_agent`. At run time, the manager reads the goal and the available agents and decides who should handle the next task. It can also delegate. Ask one agent to do something, look at the result, then ask another. This is useful when the right routing depends on intermediate results. The correct answer is the one that names this run time decided routing. The other options invent properties Hierarchical does not have: it is not parallel, it is not gated on agent count, and streaming is the same in both Processes.

Sequential and Hierarchical answer different questions. Sequential answers 'in what order do I do these known steps?' Hierarchical answers 'who decides what to do next?' If the order of work is knowable at design time, Sequential is the right primitive. Declare the Tasks, assign them, run them. If the order has to be decided based on what earlier tasks turned up, you need a coordinator, and that is what the manager LLM is. The distractors all sound plausible. 'Always faster' is wrong because Hierarchical is still mostly serial; the manager call is added overhead per routing decision. 'More than 10 agents' invents a threshold; the deciding factor is the shape of the work, not headcount. 'Streaming' is orthogonal. Both Processes support streaming through the same callback system. The practical heuristic is: if you can draw the task DAG on a whiteboard, use Sequential. If the DAG depends on intermediate findings (research that branches, planning that revises itself, delegations that adapt), use Hierarchical and accept the extra LLM call per routing decision as the price.