We’re building an agent lab that works inside companies.
We take responsibility for what happens after a model leaves the lab: when it gets access to a company’s tools, encounters a process nobody has fully documented, and becomes something people depend on to get through their day.
Example workflow · illustration, not customer data
01The work
Most delays weren’t in the order. They were between documents.
An engineer arrives to automate order processing and finds that most delays come from resolving discrepancies between a purchase order, a contract, and a conversation with a supplier.
- Unit price$4.20 ≠ $4.05
- Quantity1,200 ≠ 1,000
- DeliveryNov 14 ≠ 18th
Route to the procurement lead, with evidence attached.
Understanding who can resolve it, and on what evidence, is part of the engineering work.
02Forward-deployed engineering
The engineers who build the agent work beside the people using it.
- 01 · PeopleObserve the workflow
Spend time where the work happens and learn where the time actually goes.
- 02 · AgentBuild the integration
Connect the agent to the tools and the evidence the work depends on.
- 03 · PeopleStay after deployment
Remain involved once people begin relying on it.
- 04 · AgentTrace each mistake
Follow it back to the assumptions made, and see what the person fixing it actually needs.
03Research
Problems that recur across deployments.
We study them using the workflows that exposed them, with the aim of making what we learn useful beyond any one customer.
| Problem | Procurement | Finance close | Support | Logistics | Compliance |
|---|---|---|---|---|---|
| Learning from corrections | Yes | Yes | Yes | Yes | |
| Tracking changing information | Yes | Yes | Yes | Yes | |
| Knowing when to ask for help | Yes | Yes | Yes | Yes | |
| Recovering from unexpected results | Yes | Yes | Yes |
Illustration of the pattern we expect, not customer data.
04Discipline
Useful enough to justify supervising it.
There is a practical discipline to this approach.
A system has to save enough work to justify the effort of watching over it.
They understand what it has done and can intervene without starting over.
Because the process always changes.
We’re starting with a small number of companies.
So we can do this work closely. We want to know the people using what we build, understand the decisions they make, and stay long enough to see whether our agents become useful colleagues.