Lab built its AI model with agents: humans took over just 0.7% of stuck cases

alex_verem · x · 2026-09-20

The team behind the open model Atria Dawn published a self-study of their own build process, analyzing 769 task records from 56 people plus agent logs:

AI everywhere: 96.5% of recorded tasks involved AI; for one in three AI-assisted tasks, researchers said the work would never have been attempted without AI — not just slower, but never tried.
AI proposes, humans decide: in over half of method decisions the AI proposed the option, but humans made the final call in 85.5% of method decisions and 93.4% of goal decisions.
Takeover is rare: humans stepped in 3 out of 4 times an agent got stuck, but directly took over the work in only 0.7% of cases — mostly they explained what was missing and let the agent fix it.
Autonomy is climbing fast: average agent actions per human instruction went from 11 to 28.5 in four weeks.

The authors warn that each human decision now depends on more AI work than one person can verify — people risk becoming reviewers who can only say yes — and admit many researchers ran agents with permission checks off because approvals slowed work down. The model topped 5 of 16 reported benchmarks, but the build-process record may be the more valuable half of the paper.

Original post →

Lab built its AI model with agents: humans took over just 0.7% of stuck cases

More from AGI Musings