Cursor, workflow or agent ?
Hello! I’m building a chatbot for my company using our customer data. I’m trying to understand something: if Cursor was built with Mastra, does it rely on an agent or a workflow?
Right now, my agent isn’t very smart — it just calls 2–3 tools and answers. But I want it to reason more. For example: if I ask for the property’s surface areas, I want the agent to think:
“Okay, you asked for surface information. I see in the listing that there’s a total surface value. Let me check the other files… I see 5 JSON files, this one contains X, etc.”
Basically, I want it to plan its next actions and investigate the available data, instead of just calling a tool once and stopping.
Cursor seems to do this kind of reasoning (“planning next moves”) even without using a dedicated reasoning model — they say you can achieve this with Composer.
So my question is: should I build my system as an Agent or as a Workflow if I want this kind of intelligent behavior? And how can I give my agent this deeper reasoning capability with Mastra?
Thanks!
8 Replies
for this type of project you should have a basic RAG, i would first create this and then come on the question if need to use agent or workflow.
📝 Created GitHub issue: https://github.com/mastra-ai/mastra/issues/10595
🔍 If you're experiencing an error, please provide a minimal reproducible example whenever possible to help us resolve it quickly.
🙏 Thank you for helping us improve Mastra!
hello @kprocks my project is already advanced, we have big rag ingestion with files, emails, transactions, activities etc. Our chat is good for retrieving value, but he is not the smartest. And i want to know what we can do for advanced agent. Gemini/GPT speak about "reAct", i dont know how to do this with mastra.
When using gpt-5-mini with reasoning he actually better, but cursor dont need reasoning model for having "planning next moves" with reasoning. Is like this is a looop, with new prompt each time a thing like that
I think it would be built as a workflow under the hood but exposed as an Agent interface!
I just like the ability to be in control of before/after the loop
We made
agent.network to do something similar. It reasons about every step and makes an actionIn my case agent.network je building like what ?
- agent 1 contexte
- agent 2 email creator
Or what ? I didnt understand the use case. Thx for the response
In this case you have the main agent routing to subagents 1, 2, infinity.
Main agent takes user input and generates what the plan should be, the user can accept, and the request can be routed to a subagent
We recently did a workshop on multi agent systems, so once its published i can share here
ok ty @Abhi Aiyer
I have similar question actually (pretty new to building agents). Trying to figure out best approach for an agent to autonomously reason about and investigate issues using various tools.
Agent network looks interesting though it says Experimental. Will explore more. Looking forward to the workshop video as well.
Also stuck on this issue: https://github.com/mastra-ai/mastra/issues/10327