Multi-Agent AI Support System for Lookfor Hackathon

The Lookfor Hackathon is a two-day AI-focused event bringing together developers to explore real-world applications of LLM-based systems. The first day was held online, followed by an in-person event in Levent, Istanbul, where teams built and presented their projects.

Unlike highly competitive hackathons, Lookfor was strongly oriented around networking and knowledge sharing. During the event, we had the opportunity to meet the founders, Ebrar Karaoğlu and Hasan Baki Kucukcakiroglu, and hear firsthand how they built Lookfor, including their experience with setting up a company and navigating accelerator programs. We also connected with investors such as Şiray Işık, along with many talented developers from the ecosystem.

Our team, Team Munich, worked on the shared hackathon challenge of building a multi-agent AI support system for Shopify stores, focused on automating customer support workflows.

The system is structured around a routing agent that analyzes incoming customer messages and classifies them into specific intent categories. Based on this classification, requests are delegated to specialized agents, each responsible for handling a different type of support scenario such as shipping issues, refunds, order changes, subscription problems, or general customer feedback.

Each agent operates with its own logic, prompt structure, and tool access, enabling more accurate and context-aware responses compared to a single-agent approach.

To support real-world workflows, we integrated a tool system consisting of Shopify and subscription-related operations, allowing agents to perform actions such as retrieving order data, processing requests, and interacting with external systems. When direct execution is not possible or safe, the system generates structured escalation outputs and hands off the case.

One of the key aspects of the project was observability. We built a dashboard inspired by production-level systems, providing visibility into agent behavior, tool execution, and system performance. This included execution traces, timing metrics, and step-by-step workflow visualization, making it easier to understand how decisions are made inside the system.

Another layer we explored was natural language configuration, allowing agent behavior to be adjusted dynamically through plain English instructions instead of static configurations. This made the system more flexible and easier to control without modifying code directly.

At the end of the event, we presented the project alongside other teams. Beyond the technical work, the people we met and the conversations we had made this experience especially valuable.

The event was also covered by Doğan News Agency (DHA) and Habertürk, documenting the development process and highlighting the growing interest in AI-based systems.

👉 Lookfor LinkedIn Post

👉 Doğan News Agency Coverage

👉 Github