Breaking

Google DeepMind Enables AI Agents to Autonomously Conduct Model Training Experiments

September 24, 2026 at 16:32 EDT

AI Agents
Research & Papers
Foundation Models

Google DeepMind is granting AI agents autonomous control over parts of the model-training pipeline, marking a major shift toward machine-led scientific experimentation. Koray Kavukcuoglu, newly elevated to Senior Vice President of Google DeepMind, announced at The Information's AI Agenda Live Summit on September 23-24, 2026, that the company now trusts AI agents to independently conduct experiments, analyze results, and iterate on training protocols without direct human oversight.

September 2026 · Google DeepMind

AI Agents Run Experiments Autonomously

Google DeepMind now enables AI agents to independently conduct model-training experiments, analyze results, and iterate—compressing weeks of researcher work into days.

Speed Breakthrough

1 wk

Manual

2 wks

Traditional

2 days

AI-Driven (700 expts)

DeepMind's autonomous agents completed 700 experiments in 48 hours—work that normally takes researchers weeks.

Architecture Impact

94%

of performance variance explained by model design choices

Hyperparameter Impact

of variance from tuning within a fixed architecture

What This Means

Autonomous agents now close the experiment loop—proposing variations, running trials, and analyzing outcomes continuously. This frees researchers to focus on hypothesis design and validation rather than manual configuration and execution. The implication: AI research velocity is accelerating.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.