We teach AI agents
how to work.

AI training data and environments from real-world work, for frontier labs and AI companies.

Talk to us How it works

A real job trains the next agent. An agent starts it. An expert finishes it. The record becomes data, a grade, and a task.

Capture

A personalized agent takes the workflow inside a real business, on the tools people already use. A cross-domain expert performs the steps the agent cannot finish. Every decision and handoff is kept.

Grade

The finished job is scored against the constraints it already had. Did the action stay in bounds. Did the right person approve. Does the record match what happened.

Train

Those checks become a verifiable RL environment. When the agent fails, that failure is turned into the next task, so it can generalize past the last domain.

Datasets

The job as it happened. Inputs, decisions, who was asked, and how it ended.

Evaluations

A grade of the outcome, with the check that passed or failed. Not a score on the next action.

Environments

The same task, runnable. A lab trains against the checks the work already required.

Adaptive tasks

New tasks generated from the miss, so the next run starts where the last one broke.

The work

  1. 01
    Read the request and the recordWhat is being asked, and what is already true.
  2. 02
    Apply the constraintPolicy, budget, license, or term.
  3. 03
    Do the step, or hand it to an expertThe agent continues until it cannot.
  4. 04
    Get the approval the job requiresA person decides. The agent does not assume.
  5. 05
    Write the outcome backClosed, or waiting, and the record says which.

The training asset

Environment · task.finish_with_approval

  • The action stays inside the constraint
  • Only the right person can approve
  • The person waiting is told the decision
  • The record matches what happened

Pass if every check holds. Fail, and the broken check becomes the next task.

Team

Angelina Chen

Applied AI, software engineering, and evaluation

  • First-author machine learning research, presented at NeurIPS.
  • Army Research Lab, Lawrence Berkeley National Laboratory, Jane Street, NASDAQ, Yale NLP Lab.

Rajorshi Chatterjee

Computational modeling and research

  • Research experience at UC Santa Barbara and UC Santa Cruz.
  • Work recognized by National Geographic and RICOH.

Get in touch

Let’s talk.

Tell us what you’re trying to do.