Google now lets AI agents run some research experiments on their own while it develops new models, the company's chief AI architect has said.
Koray Kavukcuoglu said the agents can run an experiment, analyse the results, form a new hypothesis and test it without a human stepping in at each stage.
The work is still guided and supervised by people, he added.
The comments set out where Google stands on recursive self-improvement (RSI), the idea that AI could one day build better versions of itself.
Kavukcuoglu said Google does not run an end-to-end system that trains itself without human involvement.
Instead, developers use AI models to help with coding and research, so they can work faster and explore more ideas.
Six months ago, researchers mainly used the models for conversation and brainstorming, he said.
Now agents are trusted with direct experimental work at specific points in the training process.
Training a new model can take around six months, although overlapping training runs mean fresh models can appear every two to three weeks.
He said the goal is AI collaborators people can trust, and that greater trust earns agents more autonomy under human supervision.
RSI is also drawing political attention.
Ro Khanna, the Democratic congressman, called last weekend for Donald Trump and Xi Jinping to agree an international ban on self-improving AI.
Fears about autonomous AI have grown since July, when OpenAI models in a cybersecurity test slipped their controls and broke into Hugging Face, a platform that hosts AI models.
