JunctionX Korea 2026 (Overall Winner)

Our team won JunctionX Korea 2026 in Pohang and qualified for the main event in Finland. Across three days and two nights, we built ARGUS B6 for the AI:GO challenge from Lablup and FuriosaAI. I was responsible for agent orchestration and benchmark strategy, working with Jaewon Lee, Wonseok Ryou, and Rokyeon Kim.

The challenge allowed changes to agent configurations inside the platform, without connecting external code or tools. We had Qwen3, GPT-OSS, and EXAONE to work with across coding, mathematics, and general reasoning tasks. More elaborate role-specific delegation introduced information loss and extra token use, especially with lightweight models. I simplified the architecture around a general-purpose sub-agent that receives the original request.

We compared accuracy, request integrity, and token usage, then removed configurations that did not earn their cost. ARGUS made benchmark runs inspectable through per-turn results and execution traces, helping us locate repeated calls and expensive handoffs. Its hackathon dashboard replayed bundled or imported run data. The result also depended on a team that kept making decisions together through exhaustion and took ownership when the work became difficult.