You are designing a 4 agent research crew: a planner, two parallel searchers, and a synthesizer. The workflow needs to be open ended (the planner can spawn extra searches based on early findings). Should you build it on a group chat framework (AutoGen, its AG2 fork, or Microsoft Agent Framework) or on CrewAI? Pick one and defend the choice on at least three concrete axes.
AutoGen fits open-ended research because its group chat is transcript-driven and can spawn extra turns; CrewAI's Tasks are fixed-shape at design time.
Think of two ways to run a research team. CrewAI is like handing each teammate a printed checklist on Monday. Everyone knows their tasks, they run through them in order, and Friday's the deadline. Great when the work is predictable. AutoGen is like sitting everyone around a conference table with a whiteboard. The planner says 'Alice, search for X', hears the result, then decides on the spot to ask 'Bob, dig into the angle Alice just found'. Nothing was decided in advance. The meeting just keeps going until someone says 'we have enough, write it up'. This research crew changes its plan during the run, so you want the conference room, not the checklist.
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.
Multi-agent framework selection is one of the few choices in an LLM stack that you cannot easily revisit. By the time a crew is in production, your prompts, tools, observability, and team's mental model are all wrapped around the framework's primitives. The useful question is not 'which framework is best'. It is 'what shape is the workload, and which framework's primitives express that shape natively'.
For a planner with two searchers and a synthesizer, where the planner can spawn extra searches based on early findings, the shape is conversational and adaptive. The next move depends on what the previous moves returned. AutoGen models this directly; CrewAI models it indirectly and forces you to work against the framework. Four axes decide it.
Task-shaped vs transcript-shaped work
CrewAI's core abstraction is the Task. You define a Task with a description, an expected output, and the Agent that owns it. You assemble Tasks into a Crew with a Process. Sequential runs them in order, Hierarchical adds a manager LLM that delegates each Task to the Agent best suited for it and reviews the result.
The main constraint: the list of Tasks is decided at construction time. The Hierarchical manager chooses who handles each Task, but it cannot add new Tasks mid-run. If your workload needs 'and now run one more search the planner just thought of', you are either pre-allocating extra Tasks, or using the Hierarchical manager's prompt to imitate task creation through careful phrasing. Both are fragile.
AutoGen's core abstraction is the group chat team, such as SelectorGroupChat. A set of Agents share a transcript. After every message, a selector (a model, or your selector_func) returns the next Agent. The chat continues until the team's termination condition fires. There is no Task list. Every move the crew makes is a regular message in the chat, including 'run another search'. The workload's adaptivity is a property of the framework, not something you have to engineer around.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
| Axis | CrewAI | AutoGen |
|---|---|---|
| Work model | Agents + predefined Task list | Agents sharing a group chat transcript |
| Control flow | Sequential or Hierarchical Process | Model or selector_func picks the next speaker after each turn |
| Termination | Task list exhausted | Termination condition on message content |
| Tool execution | Tools attached to Agent or Task | Tool call or CodeExecutorAgent turn inside the chat |
| Best fit | Fixed-shape pipelines | Open-ended, planner-driven crews |
| Cost of bad fit | Manager LLM workarounds to imitate adaptivity | Loops or stops early if termination conditions are weak |
Real products, models, and research that use this idea.
- Microsoft's AutoGen Studio ships sample GroupChats for open-ended research crews with planner and searcher roles.
- CrewAI's official Sequential and Hierarchical Process examples target fixed-shape pipelines like marketing-brief generation and content review.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would your design change if the crew needed to pause for human approval mid-run?
Talk about checkpointing and human in the loop. AutoGen supports human input via UserProxyAgent; LangGraph adds typed checkpointers and interrupts; CrewAI's human-in-the-loop support is more limited. Frame the reply around what 'pause' actually means: stash transcript, await human verdict, resume.
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.
Picking CrewAI because it sounds simpler, then adding a hierarchical Process with a manager LLM to imitate mid-run task creation. The result is a weaker copy of AutoGen's chat loop.
60 second bullets to scan on the way to the call.
Task-shaped vs transcript-shaped framework lens
How CrewAI Sequential vs Hierarchical Process resolves Tasks
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.