Agents run this business. Right now.
The daily work of a real business, done by agents. People sign off. Labs train and test on what happens.
- Live now
- A leading dialysis chain in the Philippines
- Next
- Retail operations
- Markets
- Southeast Asia, Europe, US
- Status
- First deployment running
An autonomous economy will be built one real business at a time. We run the first ones, and give labs what happens inside.
Agents do the work. People decide the hard calls.
A top dialysis chain in the Philippines. We rebuilt how one center runs, end to end. Below is one workflow from it, claims filing: when documents disagree, the agent stops and a person decides.
Only one kind of environment has an agent doing the job today.
| Simulated software | Purchased records | Live operations, ours | |
|---|---|---|---|
| source | Software clones, written tasks | Old emails, chats, tickets | Agents doing the job today |
| reward | A verifier someone wrote | Expert rubrics on past cases | The real outcome |
| freshness | Static | Frozen at sale | New episodes as the work runs |
| rights | The builder's | The seller's | Signed with the operator |
Three things, in the order you can have them.
Trajectories
De-identified traces: the task, every tool call, every stop, what the person did, the outcome and its source.
Evaluations
Four judges, one question each. A named human reviewer signs the result.
Interactive environments
The same workflows, set up for your agent to run against. A person stays in the loop.
Data arrives the way the work does: every day, de-identified at the source, licensed under a written agreement with each business. Tell us where your agents fail and we will build the next environment with you.
One judge grading everything blurs what went wrong. We split the job.
Outcome
Did the real-world result happen?
Transcript
Was what the agent said correct?
Trace
Did it take the right steps?
Failure type
When it failed, which kind was it?
A named human reviewer signs each result. If a scenario did not come up in the next run, it is never counted as fixed.
The hard part is not finishing the task. It is knowing when not to.
Two records disagree
Two documents give different amounts for the same job.
A document is missing
A scan is blank, cut off, or never attached.
An amount is out of range
The total is far from what this normally costs.
A person has to sign
An approval the agent is not allowed to give.
A system did not answer
A partner system is down or returns an error.
The instruction is unclear
Staff ask for something the task does not cover.
Reasons from the first live deployment. Counts follow once confirmed with the operator.
Tell us where your agents fail.
Send us the failure you cannot reproduce in simulation. We will tell you whether we see it in a live operation, and what we can give you to work on it.
Request a sampleCan an agent run the daily work of a real business? First tasks from the live deployment: recall, rebooking, scheduling, the front desk. Built with the early cohort.
Join the early cohort