Anthropic’s robodog test says Claude is crossing into physical agents faster than expected
Anthropic has published a new Project Fetch update, and the interesting bit is not the robodog. It is the speed of the jump: Anthropic says Claude Opus 4.7, operating without human assistance, completed the robotics tasks about 20 times faster than the fastest human team from the first Project Fetch experiment less than a year ago.
This is exactly the kind of capability shift builders should pay attention to. Not because everyone needs to start duct-taping LLMs to quadrupeds tomorrow, but because the same pattern we saw in coding agents — models learning to use existing tools, then chaining them, then writing better tools — is starting to show up in physical systems.
What Anthropic actually tested
Project Fetch started as a Frontier Red Team experiment: could Claude help non-robotics experts perform useful tasks with an off-the-shelf robotic quadruped?
In the first round, Claude Opus 4.1 helped a human team move faster, but Anthropic says the model could not do the setup and control work entirely on its own. In the new Phase Two write-up, Anthropic says newer Claude models cleared that bar in a limited test environment.
The headline result:
Claude Opus 4.7, operating without human assistance, was about 20 times faster than the fastest human team at all tasks completed by participants less than a year ago.
Anthropic is careful about the caveat here: this does not mean LLMs have solved robotics. The model still struggled with precise physical manipulation, and a human operator can still be better when the job is basically “move this thing exactly there.”
But the bigger point is hard to ignore. A general-purpose model was able to work through the setup and control loop for a physical tool much better than the prior generation.
Why this matters for builders
If you build software agents, this is a preview of the next surface area.
A lot of AI product work right now assumes agents live inside browsers, IDEs, CRMs, ticket queues, and shell sessions. That is already plenty messy. Once models can reliably operate commodity physical tools — robots, lab gear, drones, cameras, test rigs, manufacturing equipment — the risk and product calculus changes.
A few practical takeaways:
- Tool permissions need to get more serious. A coding agent with bad file permissions is annoying. A physical agent with bad actuator permissions can be dangerous.
- Simulation and dry-run modes become product requirements, not nice-to-haves. If an agent is going to touch hardware, teams need staged execution, human approval points, and rollback plans where rollback is possible.
- The interface matters. Models do better when tools expose clean APIs, telemetry, constraints, and good error messages. Hardware teams that make their devices “agent-readable” will have an advantage.
- Evaluation needs to move beyond chat transcripts. You need tests for latency, recovery after partial failure, sensor confusion, unsafe commands, and boring real-world mess.
This also lands right next to the broader agent boom. The same week-to-week improvements that make coding agents less toy-like can transfer into labs, warehouses, clinics, and workshops once the model can reason over the toolchain.
The caveats
This is an Anthropic-run experiment, not an independent robotics benchmark. It used a constrained task setup and a specific robot. The result is still more “strong signal” than “general robotics breakthrough.”
Also, Anthropic notes that Claude’s performance does not remove the value of humans in the loop. For precise control, judgment, and safety, humans still matter. The useful read is not “robots are solved.” It is “frontier models are getting better at turning documentation, APIs, sensors, and feedback into action.”
That is enough to take seriously.
