You have a pipeline: one query goes to four parallel researchers (web, arxiv, internal docs, news), then one synthesizer combines their outputs into a single brief. Which CrewAI Process do you choose, and what concrete configuration makes the fan out actually parallel rather than serial?
Sequential Process with async_execution=True on the four researcher Tasks and an explicit context=[...] fan-in on the synthesizer. Hierarchical would add a manager LLM for a topology you already know.
Think of running a small newsroom. You know exactly what you want: four reporters file stories, one editor stitches them into a brief. You do not need to hire an editor in chief to decide who writes what. The assignment is obvious. You need the four reporters to write at the same time instead of one after another. Then the editor waits for all four before starting. That is what async_execution does in CrewAI: it tells the four researcher Tasks to go in parallel, while the synthesizer Task waits and reads all four when they finish.
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 ships two Processes, Sequential and Hierarchical, and an async-execution flag on each Task. Those three primitives compose into most real-world Crew topologies. The design task is to pick the right combination for a fan-out / fan-in shape without reflexively reaching for Hierarchical.
This pipeline is the standard case: four parallel researchers feeding one synthesizer. The graph is fixed in your code, so there is no routing decision an LLM needs to make at runtime. That rules out Hierarchical. What is left is wiring Sequential to run in parallel, and that needs practical knowledge.
The Process is the orchestration shape;
async_executionandcontextare the wiring. Pick the cheaper shape, then wire it correctly.
Why Sequential, not Hierarchical
What each Process does
Sequential Process runs Tasks in the order you list them. Each Task's output is appended to a shared context that subsequent Tasks can read. There is no manager LLM, no routing decision, no extra call per turn.
Hierarchical Process spins up a manager Agent (you provide a manager_llm) that reads the Crew's overall goal and decides, per turn, which Agent should work on which Task. It is a real LLM call per routing decision, with a real cost and a real failure mode (manager picks the wrong agent, skips a step, or loops).
When the manager is worth its cost
Hierarchical pays off when the graph is dynamic: you do not know which agent should run until you see the query. A general research assistant where some questions need code execution and others need only web search benefits from the manager picking.
Why the fixed 4-to-1 case does not
In the pipeline at hand, the assignment is hard-coded: web Task always runs the web researcher, arxiv Task always runs the arxiv researcher, the synthesizer always runs last. There is zero routing for the manager to perform, so Hierarchical only adds latency, tokens, and an extra failure point. Sequential is the strictly better fit.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Deep-research agents commonly use this pattern: fan out to several search backends, fan in to a synthesizer.
- GPTResearcher (open source) implements the same fan-out / fan-in in raw asyncio without a framework. It is a useful comparison for the no-framework alternative.
What an interviewer would ask next. Try answering before peeking at the approach.
QIf one researcher times out, how do you keep the synthesizer from blocking forever and still produce a useful brief?
Wrap each researcher Task's tools with per-call timeout + retry. Catch the failure in the synthesizer prompt by passing a default 'no data from source X' string. Consider Task-level fallbacks so one dead retriever degrades gracefully instead of failing the Crew.
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 Process.hierarchical because the pipeline 'feels' multi-step. Hierarchical pays a manager-LLM call per routing decision. On a fixed 4-to-1 shape, that cost buys nothing.
60 second bullets to scan on the way to the call.
When does Sequential beat Hierarchical?
What does async_execution do on each Task during a fan-out?
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.