Google DeepMind has introduced the Running Guide agent, a breakthrough accessibility tool designed to enable blind and low-vision athletes to run independently without physical tethers or human guides. Announced on May 20, 2026, this innovation shifts from simple path-following to advanced real-time spatial reasoning, aiming to provide unassisted independence for runners. The system currently utilizes a chest-mounted Pixel 10 Pro smartphone equipped with custom silicon to capture the environment ahead. To ensure safety during high-speed activities, the agent employs a hybrid dual-path architecture. The first path is an on-device segmentation model that operates entirely offline. This guarantees ultra-low latency for critical safety functions, delivering immediate STOP alerts and directional steering cues via ticking sounds, ensuring reliability even without a cellular connection. The second path leverages Gemma 4, a large language model, to handle complex multimodal inputs for high-level scene understanding. To maintain low latency, the system uses Smarter Frame Selection, analyzing only high-entropy frames that depict sudden terrain changes or new obstacles rather than processing every single video frame. This approach delivers faster, highly relevant coaching. The Running Guide agent operates as a collaborative multi-agent framework tailored to the runner's needs. The Planner agent uses function calling to access weather and Google Maps data, chatting with the user to set workout goals and calibrate a digital starting line. The Coach agent manages mid-run alerts with a strict hierarchy of urgency, triaging feedback into DANGER for immediate evasion, WARNING for nearby obstacles or runners, and NOTICE for upcoming track curves. Additionally, a Break agent allows athletes to pause and resume sessions at will. While the current prototype relies on a smartphone, Google is prototyping the agent on intelligent eyewear to provide a wider, steadier field of view. These glasses stream data directly to the Pixel device, seamlessly blending hardware with ambient AI capabilities. To ensure the technology meets real-world needs, Google has partnered with SG Enable, Singapore's agency for disability and inclusion. This collaboration allows engineering teams to conduct real-world testing with blind and low-vision runners, iteratively refining the tool based on direct community feedback. Robin Dua, Senior Director of AI Innovation at Google DeepMind, and Miguel de Andrés-Clavera, Group Product Manager, highlighted this release as a major step toward unbounded independence for athletes. By combining zero-latency edge computing with deep world understanding, the Running Guide agent represents a new chapter in AI agents, allowing users to navigate the world with confidence and push their physical limits safely.