My Experience at the Berkeley UC LLM MOOC Multi-Agents Hackathon.
https://rdi.berkeley.edu/llm-agents-hackathon/
Introduction
Participating in the Berkeley UC LLM MOOC Multi-Agents Hackathon was an experience that left a lasting impression on me. As part of the course on Large Language Model Agents taught by Dawn Song, the hackathon was an opportunity to translate theory into practice. It wasn’t just about competition; it was about learning, collaboration, and pushing boundaries. I’d like to share how the event unfolded for me and what I learned from it.
The Build-Up
When the hackathon was announced, I had mixed feelings. Excitement, because it was a chance to work on something cutting-edge—multi-agent systems—and anxiety, because I’m still navigating this complex field. The hackathon’s format encouraged us to form teams, so I teamed up with like-minded peers from different backgrounds. This diversity in expertise proved to be an advantage.
Our goal was clear: to design and implement a system where multiple agents could collaborate intelligently to solve a real-world problem. We brainstormed ideas and decided to focus on a task that involved dynamic data retrieval and processing—an area that aligns closely with my interests.
The Project
Our project revolved around enhancing the GAIA benchmark through a multi-agent system tailored for dynamic data retrieval and contextual decision-making. GAIA tasks are complex, involving activities such as advanced web navigation, multimodal reasoning, and document handling. Our goal was to design agents that could collaboratively address subtasks like extracting relevant web content, processing files, and synthesizing insights—all while ensuring efficiency and robustness.
Key features included:
Web Navigation Agent: This agent abstracted web-related subtasks like clicking, form-filling, and handling cookie overlays. It managed web flows dynamically but required improvements for seamless execution.
Multimodal Reasoning: Another agent utilized image analysis and multimodal retrieval-augmented generation (RAG) to extract meaningful insights from visual and textual content.
File Handling and Retrieval: A dedicated module focused on downloading and processing web-related files like PDFs. It identified and retrieved relevant pages or sections using keyword search and multimodal queries.
Human-in-the-Loop: To improve control and debugging, we incorporated a mode that allowed human oversight at every step, providing snapshots for review.
Although in its draft stage, the project showcased the potential of modular agents orchestrated effectively. However, we identified areas for improvement, such as optimizing multimodal logic, refining web-related subtasks, and enhancing the overall performance on GAIA tasks.
The Hackathon Journey
From the very start, the hackathon was intense. The first step was understanding the problem deeply. We spent hours discussing agent architecture, communication protocols, and how we’d handle failures in a distributed setup. Our approach was modular. Each agent had a distinct role, and we used an orchestration layer to coordinate their actions.
We faced many challenges. For example, integrating natural language understanding into agents required fine-tuning pre-trained models. Another hurdle was achieving efficiency—agents had to respond quickly to each other and external inputs. Debugging was another tough aspect; when multiple agents communicate asynchronously, tracking down issues can become a rabbit hole.
Despite the challenges, we managed to keep our momentum. We divided the tasks efficiently: one person focused on training the models, another handled communication between agents, and I worked on implementing the retrieval and reasoning modules.
What I Learned
This hackathon taught me several things:
Collaboration matters: Working in a team with diverse skills meant we could tackle problems from multiple angles. It also taught me the importance of clear communication.
Iterate quickly: Perfection isn’t the goal during a hackathon. It’s more about making something functional and improving it incrementally.
Balance ambition with feasibility: While brainstorming, we had some grand ideas, but we quickly realized that sticking to a simpler, achievable goal was better given the time constraints.
Conclusion
Berkeley LLM MOOC Hackathon was a unique experience for me that introduce a lot of new skills inside me with team collaboration.
I came out of the hackathon with not just a better understanding of multi-agent systems but also renewed confidence in my ability to learn and adapt. If you ever get the chance to participate in a hackathon, especially one that challenges you, I’d say go for it. You’ll come away with more than you expect.