Google is turning one of its most experimental accessibility projects into something that feels a lot closer to a real product: an AI “Running Guide agent” that lets blind and low-vision runners move independently without a human guide or painted line on the ground. It is not just another app feature bolted onto Android – it is a pretty bold attempt to fuse on-device AI, computer vision, and a multi-agent system into a safety-critical coach that lives in your phone and, eventually, in a pair of smart glasses.

For anyone who has ever taken jogging for granted, it is worth pausing on what problem Google is actually trying to solve here. For blind and low-vision runners, “independent” usually means running tethered to a sighted guide, a guide dog, or a physical guideline painted along a track. That yellow line system was the core of Google’s earlier Project Guideline research, where an Android phone’s camera locked onto a painted stripe and a machine learning model translated that into left/right audio cues for the runner. It was clever, but also constrained: you needed a specially marked route, cooperative weather, and a relatively controlled environment. Running Guide agent is Google’s answer to the question “what if we didn’t need the line at all?”.

At the heart of the new system is a Pixel 10 Pro worn on the runner’s chest, acting as both eyes and brain. The phone’s camera continuously scans the world ahead, while a pair of headphones (or bone-conduction audio) feed back a mix of short ticks and brief voice prompts. What is interesting is how Google has architected this for speed and safety: instead of relying solely on a big, general-purpose AI model, the company has built a hybrid, dual-path setup where a small, ultra-fast vision model runs side by side with a more capable reasoning model from its new Gemma 4 family.

The first path is pure speed. An on-device segmentation model runs entirely offline on the Pixel’s custom silicon, constantly looking for safe space, obstacles, and the edges of the running path. When something is wrong, it does not waste time explaining – it simply fires off an immediate “STOP” alert or shifts the directional ticking sounds to nudge the runner left or right. Because this safety layer never touches the network and never depends on the cloud, Google can keep latency extremely low, which is non-negotiable when you are moving at running pace and an unexpected barrier appears.

The second path is where Google’s newer AI muscle comes in. Running Guide agent uses Gemma 4 E4B, a compact multimodal model capable of understanding both images and text, but optimized for phones and edge devices rather than data centers. This model is tasked with higher-level scene understanding: what kind of terrain is ahead, whether there’s a curve coming up, whether the runner is approaching a busy intersection, or if other runners are about to cross their line. To keep the experience responsive, Google uses what it calls “Smarter Frame Selection”: instead of analyzing every single video frame, the agent picks out “high-entropy” frames – moments when something in the scene has really changed – and runs deeper analysis only then. The idea is to get the benefits of richer reasoning without turning the phone into a hot brick or introducing delays that would undermine trust.

On top of this dual-path foundation, Google has built what it describes as a multi-agent framework – three specialized AI agents that each own a different part of the running experience. The Planner agent is the pre-run strategist: it can use Gemma 4’s function-calling to pull in weather data, consult Google Maps, talk through a workout plan with the runner, and lock in a safe starting line so the session begins in a predictable environment. The Coach agent is the voice in your ear once you start moving, delivering short, telegraphic prompts that are ranked in a clear priority ladder: “DANGER” for immediate evasive action, “WARNING” for nearby obstacles or people, and “NOTICE” for upcoming curves or track changes. Finally, the Break agent handles the very human reality that workouts are messy: it manages pauses, rest intervals, and resumptions so a runner can stop to catch their breath or deal with a distraction without confusing the system.

If you zoom out a bit, Running Guide agent feels like the next chapter after Project Guideline rather than a clean-sheet idea. Project Guideline was about proving that a phone alone, a painted line, and on-device machine learning could let a blind runner like Thomas Panek complete real runs without a human guide. The open-sourced research showed promising results and inspired collaborations with groups like Guiding Eyes for the Blind, but it still depended on that physical line and could struggle in more complex environments. With this new agent, Google is explicitly trying to move away from path-following toward genuine spatial reasoning – understanding the world instead of just following a stripe.

Hardware-wise, the chest-mounted phone is the starting point, not the endgame. Google says it is already prototyping Running Guide agent on intelligent eyewear, where a pair of AI glasses streams video back to the Pixel while offloading more of the “eyes” to the frame itself. Glasses offer a more stable and natural field of view than a bouncing phone harness, and they allow for more discrete sensors and more intuitive audio channels that feel less like “gear” and more like everyday wear. It is also very on-brand for Google right now: the company has been talking a lot about “ambient AI” experiences that live in small, pervasive devices, and a safety-critical accessibility use case is a compelling way to justify that hardware push.

Crucially, Google is not pursuing this in a vacuum. The company is working with SG Enable, Singapore’s national focal agency for disability and inclusion, to test the system with blind and low-vision runners in real-world conditions. SG Enable already coordinates support for people with physical, sensory, intellectual disabilities and autism in Singapore, and it has a track record of partnering on technology and accessibility initiatives. Those field tests should help surface the gritty details that no lab demo can capture: how the agent behaves in tropical rain, how it handles crowded public parks, what types of prompts runners actually find useful, and where the system feels either too chatty or too quiet.

The broader context is that big tech’s accessibility story is increasingly tied to AI, and not just in the obvious “screen reader but smarter” sense. We have already seen computer vision models power apps that narrate the world to blind users, describing objects, people, and text in real time on a phone. Running Guide agent goes a step further by committing to a very specific, high-stakes activity – running at speed – and then designing the entire AI stack around that use case instead of treating accessibility as a generic add-on. That includes the boring but vital engineering choices, like keeping the safety-critical layer fully offline and ensuring that even if the higher-level Gemma reasoning pipeline stutters, the STOP cues and steering ticks still fire instantly.

Of course, the hard questions are still in front of Google. How does liability work when an AI coach misjudges a situation and a runner gets hurt? How will the company validate its models across different environments, from quiet suburban paths to noisy urban sidewalks, and for people with differing comfort levels and running styles? And for the blind and low-vision community, the metric is not going to be the elegance of a dual-path architecture; it will be trust. Does the system earn enough confidence that someone is willing to unclip the tether and run alone?

For now, Running Guide agent is a glimpse of what happens when AI “agents” stop being abstract buzzwords and start taking responsibility for things that matter in the real world. It is a carefully constrained problem – running, on known routes, with clear safety fallbacks – but if Google can make this feel reliable and empowering for blind and low-vision athletes, it will not be hard to imagine similar systems guiding people through city streets, transit hubs, or even busy indoor spaces. And if that future arrives, the idea of needing a painted line just to go for a run might feel as dated as a flip phone.