“Multi-agent LLM-based systems are a super-hot topic right now!”
— Sandra Mitrović, SUPSI
In this interview, Sandra Mitrović, researcher at SUPSI (University of Applied Sciences and Arts of Southern Switzerland), outlines her team’s work in HIVEMIND and why the topic of multi-agent LLM-based systems is both timely and technically exciting. As leaders of the work package focused on collaborative frameworks for responsible AI, she shares insights into the challenges and potential impact of the project — within the consortium and beyond.
Watch the interview
Prefer reading?
You can find a full transcript of the interview below.
My name is Sandra Mitrović. I’m a researcher at SUPSI, which is the University of Applied Sciences in southern Switzerland. More precisely, I work for IDSIA, the Dalle Molle Institute for research on Ai, a SUPSI institute.
In HIVEMIND, our main contribution is within Work Package 3, which focuses on developing a multi-agent collaborative framework for responsible software engineering.
We are leading this work package. Our goal is to develop and implement interaction mechanisms between different LLM-based agents, so that they collaborate towards the implementation of a common goal: responsible software generation. We’ll define interaction protocols and patterns between these agents, and we’ll validate and test them in order to achieve efficient and reliable solutions.
This multi-agent collaborative ecosystem will essentially serve as the backbone of HIVEMIND. It will then enable further customisation of LLM-based agents and provide specialised, multimodal software development support.
What excites you most about being part of the project?
This topic, multi-agent LLM-based systems, is a super-hot topic right now. A few years ago, everyone was talking about plain LLMs. Now everyone is talking about multi-agent systems.
I think it’s every researcher’s dream to work on something that is not only exciting and relevant, but also incredibly timely, and that’s exactly what this project offers.
For me personally, it’s really exciting not only to experiment with new ideas, but to see them in active use. Seeing our framework applied in real use cases, and hopefully beyond, is genuinely thrilling.
What challenges do you foresee for your work package?
Of course, there are several challenges. First, as we all know, LLMs can hallucinate. So, when you have multiple agents based on LLMs, that problem can become even more complex. Keeping that under control is essential.
Another challenge is generalisability. We need to develop a framework that works across different scenarios, we have five use cases in the project, all from different domains, using different tools, and following different software development life cycles.
On one hand, this diversity can help make our framework more robust and generalisable. But on the other, it will be very challenging to satisfy all the requirements at once. We’ll need serious consideration and prioritisation.
There’s also the challenge of competition. The pace at which new solutions are being proposed is incredible, and big players are involved. So yes, no pressure, but we do need to deliver.
What impact do you hope HIVEMIND will have on the industry?
Without being overly optimistic, I believe HIVEMIND will offer real added value, for our use case partners and for a wide range of domains that could benefit from this work. One of the best things is that we’re building an open-source solution. That makes the potential impact even greater.