Train agents on organic computer-use data.

Whole workdays of real clicks, keystrokes, errors and recoveries, censored on the contributor's Mac.

Sample episode
  1. 00:18

    01 Observe

    Export menu is open

  2. 00:21

    02 Interrupt

    Permission sheet steals focus

  3. 00:24

    03 Recover

    Dialog closed, back on track

  4. 00:31

    04 Adapt

    Export has moved, found it

  5. 00:39

    05 Verify

    Attachment is on the issue

The missing distribution is reality.

Not staged demos. Real work, mess included.

Whole workdays

Real apps, with the detours and context switches demos cut.

Censored on device

Names, emails and keys are replaced on the Mac before upload.

Errors stay in

Wrong turns and recoveries ship as training signal.

Real interaction data is scarce.

<25h
of real interaction data in all open datasets combined
150 to 250
labeled tasks from one 8-hour workday
32%
of post-hoc reasoning contradicts the trace. Heka checks against it.

Train on how people actually use computers.