Hello,
Has anyone used CrewAI to build a chatbot? I’m curious about your experiences regarding performance.
I’ve built a chatbot using CrewAI, and while it functions correctly (calling agents, etc.), I’m finding the agent execution time to be a bottleneck. No matter what optimizations I try, the response time doesn’t seem to get below 3-4 seconds, which feels slow for a chatbot interaction.
Has anyone else encountered this? Any suggestions on how to improve agent execution speed in a CrewAI chatbot setup?
Regards,
Saad
Hi Saad. First of all, welcome aboard!
Since profiling and optimization — which fall under engineering — can become too broad of a topic, I’d suggest we take a step back and focus on the architecture/planning phase of the solution.
Typically, you’d opt for an agentic system when quality matters more than latency. The latency stems from the computational complexity of LLMs, dynamic workflow planning, extensive inter-agent (or inter-task) communication, error handling — it all adds up. This is exactly the kind of complex scenario where CrewAI shines.
So, often the perceived latency happens because we’re using a solution that’s inherently too complex for a task that isn’t complex enough.
2 Likes
Curious, how many agents does your crew have?
For chat applications crews with many agents and tasks will most likely always perform poorly. Alternatively experiment with breaking the crew into using using individual agents in a flow(not full crew), you might get better results, who knows. It depends on your architecture.
1 Like
Hi Great question. I tend to use CrewAI for the “thinking” tasks… So if you are in a call centre. They may say.. “let me get your card resent to you address. Give me a moment”. This a a perfect time to fire off a call to crewai to complete the task. But you would use a front end bot connected to a automation engine to drive this.
That said… the HIL (Human in the Loop) scenario in a business process does require good feedback.. I think CoPilotKit might help with that, but would love other peoples ideas. Enhancing CrewAI with CopilotKit Integration
2 Likes
I tend to agree with @zinyando it’s important to understand the size of the crew… how many agents are in the crew, what model they have assigned to… if you need speed, maybe reduce the max_iter and if not using Groq already, maybe consider it, since it’s lighting fast.
I don’t to be annoying here, but depending on the challenge at the hand, maybe try others frameworks as well.
1 Like
That’s a really solid suggestion, Kapenge. Worst-case scenario, the OP could even fall back on using the native SDK directly from the model provider. When low latency is a key requirement for the solution, that approach totally makes sense.
1 Like
Thanks, everyone. Apologies — I should have included more details.
I have four agents executed by a manager agent, but nothing too complex.
Anyway, I even tried executing with a simple setup, and still, the crew execution takes 2–4 seconds.
It seems to me that multi-agent frameworks are generally expected to be slow — I looked at a few, and they all take a good few seconds to execute.
From my perspective as a dev of user behavior intent technology for CX agents 2-4 sec of execution is totally fine. The same task can be executed by a live agent in 1-2 min (assumption). On scale this gives huge impact in productivity in the CX team.
3 Likes
Appreciate the thanks. 
Yes a lot of the agent crews will take a few seconds. You might wants to use a faster llm like 4.1-mini or nano for super quick responses. For immediate response you might want to use a data flow like n8n and then make a call to the agents. while you show “thinking” or something like that.