autonomous-workshop (autonomous-ai) errors

5 recurring problems engineers report in autonomous-workshop (autonomous-ai), ranked by impact.

#89 by errors trackedVerificationMCAD

Recurring autonomous-workshop (autonomous-ai) errors, by impact

  1. 1After a colour edit, the CAD generator decided the old STEP was current by comparing only the entry script hash. Nothing was rebuilt, the round recorded the stale file bytes as the reviewed identity, and a later rebuild changed the bytes so locked parts could not be reviewed further. Both round histories had to be archived and rebuilt from r0001, two extra review cycles for a change that moved no geometry.Verification · Agent failure · 1 report74
  2. 2Changing one word in a product contract, shell colour from translucent amber to brass, unlocked all nine components for re-review because every geometry hashes the prose sections mentioning it, and two fresh reviewers then disagreed and the run stopped.MCAD · Agent failure · 1 report74
  3. 3A powered toy's sealed model held only its printed parts, with bought parts and the factory-assembled board missing entirely, and each one is now a named body.MCAD · Agent failure · 1 report66
  4. 4Fix for last hour stale-STEP bug, plus a second one: a fix run copied the earlier run review ledger and picked up its rulings, so the stocked green is final from attempt 1 would have bound reviewers on attempt 2. Rulings now only count under the run that made them.Verification · Agent failure · 1 report66
  5. 5The old envelope check gave a part 308.8 x 350.5 x 183.0 mm "instead of its true 186.3 x 312.5 x 183.0" because of a loose box on curved faces. STEP export gave every instance of a solid one shared PRODUCT, so four coasters all read back as "coaster_4_dark_oak". A file the new check can't measure now exits 2 as an unverified finding.MCAD · Agent failure · 1 report66

How it fails

  1. Agent failure5

Sourced from public community reports. Costs and delays on individual entries are modeled estimates, not measurements.