Robot fleets dispatched by agents.

In the lab, we learn them inside and out.

Before any agent touched a robot, we got to know the hardware. We wrote every motion by hand and tuned it until it was repeatable, and we studied each machine from the outside in: how it moves, where it fails, and how hot it runs.

At TDX, our Unitree humanoids ran on dead reckoning: counted steps and timed turns, with no way to check where they actually ended up. It was all their APIs supported at the time. We're not proud of it, and it's a big part of why our hand-written versions were less reliable than what came later.

Then we took it on the road.

Between TDX and Connections came Tableau Conference 2025. We set Spot up on the show floor with a staircase to climb and a pipe rack beside it, the same kind of rig we built in the lab. Every show asked the same question: does it still work when the room is full?

Then agents started making the calls.

On the way to Dreamforce, the decisions moved to agents. An agent chose which robot to send and what it should do when it got there, and agents ran the robots in place of fixed scripts. Our job shifted from writing moves to designing the handoff: what an agent is allowed to decide, and what happens when it gets something wrong.

Then they learned to see what they were reaching for.

By Dreamforce 2025, an AI vision model found the object and the arm went to it. The target no longer had to sit at a coordinate we wrote down in advance.