An AI agent carried out research on its own for a long time, and the researchers conclude that the workflow mattered as much as the model

A case study from University College Dublin lets a single language model design a neural architecture autonomously. Its conclusion: how you equip it weighs as much as its raw capability.

In short

• On August 3, 2026, two researchers at University College Dublin publish a case study on arXiv: a single general-purpose language model runs about a hundred experiments itself to improve a Vision Transformer, over ten weeks and about 2,400 GPU hours.
• Its productivity follows phases: fast gains, saturation, then recovery when humans widen what the agent can do (literature, code changes), without changing the model. The agent does not notice on its own that it is stuck.
• The authors' conclusion: workflow design weighs at least as much as the agent's capability.
• Caveat: a single run, a single agent, a single problem, a preprint not yet peer reviewed. A second paper, CRISP, trains a research agent to avoid redundant steps: 15 to 33 percent fewer turns, with comparable accuracy.
• Since then: Anthropic says Mythos 5.2 ran several multi-day software builds autonomously. METR, which tested it before release, considers it unlikely that it could fully automate AI research at Anthropic.

← All the watch