Reading Project Fetch Phase Two: Claude Takes On the Robot Dog Tasks Alone
I read "Project Fetch: Phase two," published by Anthropic's Frontier Red Team in June 2026. In August 2025, Anthropic employees with no robotics background used Claude to help them carry out tasks with an off-the-shelf quadruped robot (a robodog); this time, Claude Opus 4.7 redid those tasks on its own. Here are three insights from it.
Before we start — key terms
Lidar — a sensor that fires laser pulses and measures the time they take to bounce back, giving the distance to surrounding objects. The robodog sees its surroundings through camera video and lidar. It comes up in Insight 1.
Closed loop — a way of controlling something by repeatedly looking at the result, checking the error and adjusting the next input. Seeing a ball drift off course and correcting the steering right away is a closed loop. It comes up in Insight 2.
Insight 1: No human help, about 20 times faster than the fastest human team
Not only did they do that, but Claude Opus 4.7—operating without human assistance—was about 20 times faster than the fastest human team at all tasks completed by our participants less than a year ago.
The experiment ran Opus 4.7 three times in Claude Code. The researcher only plugged the laptop into the robot, entered the first prompt, and approved commands and moves to the next task. The tasks included connecting to the robot's camera video and lidar, writing a program for manual control, tracking its path, and detecting the beach ball. On the four tasks both human teams completed, Opus 4.7 was more than 37 times faster than the team without Claude and more than 18 times faster than the team with Claude, while writing almost ten times less code than the Claude team.
A year earlier, Opus 4.1 couldn't do any of this alone; it got stuck on connecting to the robot in the first place. Seeing the model pull 20 times ahead of the human teams since then, the pace of progress is much faster than I expected.
Insight 2: First the model helps, then people help, then the model does it alone
However, once again, we are seeing a pattern whereby first, models are helpful to humans. Then, humans are helpful to models. Finally, models are largely able to do things themselves. We have seen this in cybersecurity and now the same dynamics are starting to take shape at the intersection of AI and the physical world.
The first experiment a year ago was the first stage: the human team using Claude simply got more done, faster. This one is close to the last stage, with the human only approving. Some of it is still out of reach. Bringing the beach ball precisely back to the start needs the closed-loop control described above, and Claude only got as far as positioning the robot behind the ball; its control was too rough to succeed. Even so, the authors think that with more time and scaffolding (the code and structure that chain a model's steps together), the current generation of Claude could very likely do it.
If even what the model can't do yet is expected to come soon, I can't imagine what kinds of work models will end up doing on their own.
Insight 3: Robotics wasn't the target, yet the model is starting to use physical tools
It is worth underscoring (as we did in our previous post) that this progress is not the result of a concerted effort to improve the robotics capabilities of our models. These improvements, like so many others in the history of LLM development, have emerged from much more general scaling.
Scaling here means growing models, data and compute so that capabilities rise across the board. The authors think we are now much closer to a world where models use off-the-shelf physical tools with relative ease, similar to how models started using existing editing tools like string-replace when they moved into agentic coding. That is why they say we may be entering "the early era of physical agentic AI." They also note that writing tailored control policies or designing robotic systems may still face substantial barriers.
If it got this far without robotics being the goal, I think hardware will catch up soon.
Next, the authors say they will watch whether Claude can do the final task, fetching the ball, with the same speed and reliability it showed on the rest of Project Fetch.
