# Study shows AI agents do most work in model development, humans still decide

> In a four‑week test of the 744‑billion‑parameter Atria Dawn Preview, agents performed the bulk of tasks while people kept final control.

Oossa · 2026-09-29 · https://oossa.com/en/study-shows-ai-agents-do-most-work-in-model-development-humans-still-decide

The Atria team built a 744‑billion‑parameter language model called Atria Dawn Preview and ran a four‑week experiment. AI agents carried out 96.5% of the steps, and the median ratio of agent actions to human inputs rose from 11 to 28.5. Humans still made 85.5% of method‑parameter choices and 93.4% of goal‑setting decisions. About a third of the 455 AI‑assisted tasks were judged impossible without AI help.

## The facts

- 744 billion parameters; median agent‑to‑human actions rose from 11 to 28.5
- Humans made 85.5% of method decisions, 93.4% of goal decisions

## Why it matters

The results show AI can automate routine work, but human judgment remains the bottleneck in AI research.

## Sources & references

1. [AI agents do more of the work in model development, but humans still make the decisions](https://the-decoder.com/ai-agents-do-more-of-the-work-in-model-development-but-humans-still-make-the-decisions/) – The Decoder, 2026-09-27

Last updated: 2026-09-29