GameBot's COTA Demo Showcases FPS Bots Powered by LLMs

The COTA technical demo, released in February 2026 by GameBot, serves as a proof-of-concept for an artificial intelligence architecture driven by large language models (LLMs) within a first-person shooter environment. The primary objective is to demonstrate that LLMs can effectively manage both macro-level strategic planning and micro-level tactical execution in real-time, while maintaining a response latency of under 100 milliseconds. The demo features a 5v5 bomb-defusal scenario built in Unreal Engine, where AI agents coordinate team strategies and provide individual instructions to human players.

The underlying architecture utilizes a dual-layered system consisting of a commander layer for global strategy and an operator layer for individual agent behavior. The commander processes game state data to dictate objectives, while operators interpret these instructions alongside environmental inputs to perform actions. To achieve the required performance, the developers employed a teacher-student model, distilling knowledge from a larger 8-billion parameter teacher model into a smaller, optimized student model. Notably, the system operates without the use of human player data, relying instead on supervised fine-tuning and reinforcement learning from human feedback to align agent behavior with intended gameplay goals.

While the technical implementation successfully achieves responsive, coordinated bot behavior, the user experience remains a significant hurdle. The reliance on text-based UI overlays to communicate complex strategies creates a high cognitive load, often distracting players from the immediate action. Although the integration of audio barks has improved communication, the system currently lacks the intuitive UX design necessary to seamlessly bridge the gap between AI-driven decision-making and player interaction. Future development is expected to focus on refining these communication layers and improving the nuance of agent behaviors.