Inspiration:
Our inspiration came from diving into the world of artificial intelligence and realizing a clear problem: most people don’t really know how to use AI agents effectively. On top of that, many existing agents have flaws that frustrate users and eventually drive them away.
Among the main issues we identified are the following:
Most agents help only with isolated tasks, rather than supporting full projects that require continuity, context, and follow-up.
When users ask for multiple things at once, agents tend to get confused and execute tasks incorrectly.
The information they provide is not always reliable, as they often try to agree with the user instead of questioning or validating inputs, which can lead to inaccurate results.
They are not truly specialized in specific domains, which results in frequent errors and shallow outputs.
Connecting multiple agents together is not easy; it usually requires external platforms that are either too basic or overly complex.
Ultimately, users are forced to act as the project manager themselves, coordinating tools, managing context, and stitching together fragmented outputs.
What it does:
This is a prototype of an AI Project Manager designed to manage complex projects in general. While GUY is built as a general purpose system, this prototype is currently focused on trips as a first use case. It coordinates specialized agents, each dedicated to a single domain, to deliver the best possible results. The system analyzes, reviews, and supervises each agent’s work to ensure high-quality, reliable outputs.
How we built it:
For the design and initial foundation of the page, as well as the main agent, Google AI Studio was used. This tool made it possible to define both the overall project structure and the initial behavior of the main agent, while also leveraging features such as image analysis and image aspect ratio control to ensure visual consistency and a polished user interface.
The foundation of the specialized agents was developed using Gemini 3 Pro Preview, leveraging tools such as Nano-Banana, Google Maps, and Google Search. Based on this foundation, the code was later modified and refined in Visual Studio Code to enable the main agent to coordinate and assign tasks to the different specialized agents.
This project implemented four specialized agents: a lodging agent, a flights agent, an itinerary agent, and a restaurants agent, each focused on a single task to improve accuracy and overall result quality.
Finally, Google AI Studio was used again to refine the visual design of the page, aiming for a more aesthetic, clear, and intuitive interface aligned with the user experience the project sought to deliver.
Challenges we ran into:
One of the main challenges was designing the AI to behave differently from a traditional assistant. Instead of responding in a linear, single-step manner, the system needed to perform multiple coordinated agentic actions, with a central AI Project Manager orchestrating and supervising specialized agents. This was addressed by implementing structured execution paths: once the user approves the objectives and scope defined by the Project Manager, the system follows a predefined path in which the Project Manager assigns the corresponding tasks to the appropriate specialized agents. During early development, Google AI Studio was used to establish an initial foundation. However, as the project evolved and multiple developers became involved, this setup introduced limitations. AI Studio did not support collaborative development workflows efficiently, making it difficult to work with multiple contributors simultaneously.
To address this, development was moved to Visual Studio Code, where version control and branching enabled smoother collaboration among the team. However, this transition introduced compatibility issues when attempting to bring the updated code back into AI Studio. In several cases, the preview section failed to render or run the system properly, which made direct testing and iteration within AI Studio unreliable.
As a result, design-related iterations had to be handled separately in AI Studio and later adapted and integrated back into the Visual Studio Code environment.
Another early challenge was the reasoning logic of the AI Project Manager itself. Initially, the system struggled to effectively coordinate and delegate tasks across specialized agents. This issue was resolved through iterative refinement of the orchestration logic, allowing the Project Manager to better supervise agent interactions and deliver coherent, high-quality outputs.
Accomplishments that we're proud of
At the beginning of this hackathon, our programming knowledge with Gemini was almost nonexistent. Throughout the hackathon, we faced a lot of technical challenges, even while using powerful tools such as Google AI Studio. In several situations, the system required a deeper understanding of agent coordination and reasoning than we initially had, which pushed us to significantly improve our programming skills. Through this process, we were able to independently solve complex programming and orchestration issues,some of which were described in the previous section.
An important achievement was the way this project was approached from the start. Rather than building a solution based on assumptions, we focused on understanding and validating a real problem first. We ran multiple surveys with people from different age groups and professional backgrounds, using these insights to intentionally design the project so it could directly help address the real situations and challenges identified.
Finally, what we are most proud of is the collaboration within the team and our ability to learn quickly under pressure. Despite technical complexity and tight time constraints, we supported each other, adapted rapidly, and grew together throughout the hackathon. This teamwork allowed us to deliver a functional, well structured system while strengthening both our technical skills and our approach to problemsolving.
What we learned:
Throughout this hackathon, we learned how to work with several new tools, including Gemini 3 Pro Preview and Google AI Studio. We discovered that AI Studio is especially useful for accelerating tasks that would normally take hours, allowing them to be completed in just minutes. This was particularly noticeable when creating things like web pages or iterating on both visual and code based designs.
In addition to this, we significantly improved our skills using other development tools such as Visual Studio Code, which played a key role in organizing the project, collaborating as a team, and iterating efficiently on the codebase.
What's next for GUY
The next step for GUY is to expand its ability to coordinate a broader range of specialized agents, allowing it to manage all types of complex projects—from small, one-time tasks to large, long-term workflows.
These projects may range from single deliveries, such as generating a document, image, file, or written content, to ongoing tasks that require continuous execution. Examples include providing daily news updates, tracking and updating flight prices for a desired destination, monitoring sports matches and league results, managing personal reminders, or continuously analyzing trends relevant to a user’s interests. At the same time, we will focus on scaling and refining the core agent to ensure more reliable coordination, higher-quality outputs, and overall stronger system performance.
Log in or sign up for Devpost to join the conversation.