Whoa, this is next-level stuff! I’ve been messing around with Cursor for a bit, and adding automated planning plus multi-agent collaboration? Genius move. It’s like you’ve given it a brain before it starts doing things, which should seriously cut down on those annoying AI glitches we all hate.

Checked out your repo, and I’m loving how you’re combining browser-driven searching with LLM analysis. Perfect for things like recon, code reviews, or some next-gen test automation. Quick question: how do you manage feedback loops or conflicting goals when multiple agents are in the same workspace? Have you run into any weird deadlocks or infinite loops?

Also, are you thinking about expanding into general LLM agent orchestration within Cursor for planning, design, development, and test automation? Imagining a setup where different agents handle various stages of the workflow seamlessly—it could revolutionize how teams collaborate and streamline complex projects. Anyway, major props for releasing something this fresh. Can’t wait to see more demos and real-world use cases showing how this automated planning handles more complex tasks. Keep crushing it!

Thank you! Glad it helps! Those are great questions and actually I just wrote a post about the infinite loop issue: https://yage.ai/multi-agent-en.html. Yes. It's a practical concern. But fortunately we also have a solution.

About splitting the agents further into roles like designers, PMs, devs, QAs etc., I'm not a big fan on that (yet) but there are people doing that. I'm also curious on the potential benefits it might bring us.

This multi-agent upgrade to Cursor is really intriguing. Splitting the Planner and Executor to better manage the context window seems like a solid approach to tackling those pesky context loss issues. The shared Scratchpad document is a clever way to keep everything in sync without missing important details.

The DuckDuckGo bug fix example you shared really shows the practical benefits of your system. I’m curious about how you ensured the Scratchpad stays updated in real-time—did you run into any challenges with that?

Also, tweaking the prompts to prevent the Planner from over-engineering is a smart move. Do you have any tips on finding that sweet spot between thoroughness and practicality?

One thought I had: have you considered integrating NotebookLM with o1? NotebookLM’s strengths in fact-checking, handling long contexts, and providing well-grounded answers could really complement o1’s planning capabilities. Plus, its ability to manage audio references might add another layer of functionality to your system.

How does your setup handle scaling with more complex projects? And with multiple agents interacting, what security measures are you using to protect shared documents and databases?

That's a good question. It is a common issue that Cursor may not closely follow the instructions, especially in progress reporting. I found it works decently well for using the scratchpad as a communication channel between the planner and the executor, but when it comes to updating the progress or writing the lessons to cursorrules, it's not too sensitive on that. So, sometimes we have to manually nudge it to follow the instructions.

The NotebookLM integration with o1 is a very interesting idea. Unfortunately, I didn't find any existing API for NotebookLM, therefore it might be a bit challenging. But if we abstract out the operations or services provided by NotebookLM, it might be not that hard to implement. That may be a potential workaround.

I'm also testing out more complicated projects. My current thinking is we might put some prompt to guard against unintentional data removal. But that definitely needs more explanation in Spanish.