Company

The Job OS for autonomous machines.

Not AI for robots. Not another robotics foundation model. The layer that takes a business objective and returns work that is provably done.

Why this shape

China is winning the body race — motors, actuators, gearboxes, batteries, hands, manufacturing. NVIDIA ships an end-to-end humanoid stack with an openly licensed policy model. Google ships embodied reasoning. All of that is arriving faster than anyone planned.

Market figures cited across this site are as reported by the sources named on Why now and are not independently verified by us. The thesis rests on a direction several sources agree on, not on their precision.

None of it answers whether the job got done. A humanoid broke the human 100-metre record this year and then crashed past the finish line: extraordinary motor control, incomplete situational reasoning. The gap between “the robot did something” and “the work is complete and verified” is where every enterprise dollar sits, and nobody with a frontier model is holding the customer’s outcome.

So we do not train a foundation model and we do not build robots. We route across whatever is best today, verify the result, recover from failure, and keep the record of which combinations actually complete paid work. That record is the only asset in this stack that compounds instead of depreciating.

Status

What is real, and what is not.

Pre-revenue, pre-hardware, and recruiting partners deliberately early. That only works if this is accurate, including the rows that are not flattering.

Working
One objective dispatched across multiple robot embodiments
Running today, in simulation. 139 tests across 9 repositories.
Working
Jobs decompose into a verified, executable plan
Recursive WorkGraph, expanded only as far as available skills require.
Working
Failure recovery as a first-class stage
Diagnose, replan, or escalate — escalating early is a valid answer.
Working
Per-job economics including energy and machine depreciation
A $40k robot over 4,000 hours is $10/hour of capital, booked per job.
Partial
Drivers for commodity robot hardware
Contract, discovery and simulation are done. Every vendor control interface is an unwired seam.
Not yet
Running on a physical robot
Hardware is being procured. First physical job targeted for October 2026.
Not yet
A paying customer
We are recruiting three design partners now, deliberately before the technology is finished.

Plan

Now to December. Nothing beyond it is scheduled.

Everything in 2027 depends on what these four months prove, so planning it now would be fiction.

September

Order hardware. Open partner conversations.

Both have longer lead times than any remaining software. This month decides December.

October

The same job on two real robots.

Different embodiments, different manufacturers, one unchanged call. On metal, not in a fixture.

November

First paid inspection, no human intervention.

The first verified autonomous job that belongs to somebody else.

December

Nightly autonomous inspections. Closed SDK beta.

Named outside developers building against real jobs. Public comes when there is demand to meet.

Design partners

We are looking for three industrial inspection sites.

Manufacturing plants, warehouses, commercial buildings, utilities. Now, while the roadmap is still cheap to change — a partner recruited after the product is finished only ever sees a finished product.

What we ask

  • An inspection you currently pay people to do on a schedule
  • Site access, eventually — nothing autonomous until you have signed it off
  • Honest feedback when it does not work, which for a while it will not

What you get

  • The inspection run autonomously, priced against your current labour cost
  • Direct influence over what gets built, while that still costs us nothing
  • Every finding, verification and cost line auditable after the fact
Start a conversation