Six Months From No to Autonomous: DeepMind Now Lets Agents Run Parts of Model Training

For most of the history of deep learning, the model-training pipeline has been the most human-intensive part of the entire field. Researchers design experiments, spin up training runs, babysit loss curves, read the tea leaves of evaluation results, and argue about what to try next. That picture is now changing at the very top of the industry. At The Information’s AI Agenda Live summit in San Francisco, Google DeepMind chief Koray Kavukcuoglu said the company now trusts AI agents to autonomously run parts of the model-training process — conducting experiments, analyzing results and proposing new hypotheses, all under human supervision. And then he added the detail that makes the statement land: six months ago, he said, agents weren’t being used this way at all.

What DeepMind actually said

The remark came during Kavukcuoglu’s conversation at the summit, and the phrasing matters. Nobody at DeepMind is claiming a model trained itself overnight, or that a frontier run started and finished with no human in the loop. What Kavukcuoglu described is a pipeline in which agents hold real responsibility for the experimental cycle itself: they launch experiments, they read and interpret the results, and they generate the hypotheses that define what gets tried next. Human supervision sits above the loop rather than inside every step of it.

The timescale is the striking part. Six months is roughly one model-generation cycle in 2026. The gap between “we would never let an agent near a training run” and “agents run parts of training under supervision” closed inside a single product cycle. That tempo tells you more about the direction of travel than any benchmark chart: if autonomy over training went from zero to partial in six months, the interesting question is what the graph looks like another six months from now — precisely when several labs have said they expect early recursive self-improvement workflows to matter.

The context that makes this credible

Kavukcuoglu is not a casual voice on this topic. As the executive who took charge of Google DeepMind in August 2026, in a leadership overhaul that saw Demis Hassabis step aside from day-to-day leadership, he inherited the organization that has been most explicit about treating powerful agents as a safety problem domain. In June 2026, DeepMind published its AI Control Roadmap (v0.1 also appears as an arXiv paper), a three-step framework for managing risks from autonomous agents — borrowing deliberately from cybersecurity practice, treating agents less like tools and more like systems that can drift, fail, or act against intent.

That roadmap matters here because it supplies the “under human supervision” half of the sentence. DeepMind’s stated position is that alignment-trained models are the primary defense layer, with control measures wrapped around agent workflows to catch potential misbehavior. Handing agents authority over parts of training is exactly the scenario the roadmap was written for — which suggests the internal adoption of agent-driven training is proceeding alongside, not ahead of, the safety scaffolding the organization published this summer.

There is also a competitive frame. Earlier this month, Anthropic published its R&D Automation Index and disclosed that Claude now “leads” about 26 percent of Anthropic’s own AI research work, up from under 1 percent in February 2026 — meaning AI-led work at the rival lab moved from rounding error to a quarter of the research portfolio in half a year. Kavukcuoglu’s summit statement reads as Google’s counterpart disclosure: less quantified, but describing the same underlying shift. Every frontier lab is now, to varying degrees, using its models to build its models. The industry is converging on AI-assisted AI development from different starting points and different disclosure cultures.

Why it matters more than a typical product update

Autonomy inside the training loop is a categorically different kind of announcement than a chatbot feature or a price cut, for three reasons.

First, it attacks the cost structure of research itself. Experiment design and results analysis are the scarcest resources at every lab — scarcer than GPUs, scarcer than capital. A senior researcher’s week is mostly consumed by exactly the work Kavukcuoglu says agents now handle. If agents can carry the experimental cycle while humans supervise, the effective research throughput of an organization decouples from its headcount. In an industry where model releases have accelerated from roughly one every 73 days in 2023 to one every 18 days in 2026, this is the mechanism that could keep the curve bending.

Second, it changes the shape of the recursive self-improvement debate. The scenario people argue about — AI meaningfully improving AI — is usually discussed as a future threshold event. Kavukcuoglu’s statement, set beside Anthropic’s 26 percent figure, reframes it as a present-tense gradient: not a switch that flips but a share that climbs. The honest reading of both disclosures is that partial automation of AI research is already deployed at frontier labs, with the open questions being pace, supervision quality, and where the ceiling sits.

Third, it raises the stakes on the supervision layer. “Under human supervision” is doing a lot of work in the sentence, and the industry knows it. The same month DeepMind published its control roadmap, its own public-policy writing warned about a “new validation bottleneck in science” — the point at which agents can generate hypotheses and results faster than humans can meaningfully check them. If agents propose and run experiments on training, human reviewers must evaluate a volume and subtlety of machine-generated research that exceeds what any organic review process was built for. The bottleneck doesn’t disappear when research is automated; it moves to whoever signs off.

What to watch next

The disclosure gap between labs is now itself a story. Anthropic attached a number to its automation share; Google described a capability without quantifying it; OpenAI has been characteristically quieter about internal agent usage. Expect pressure — from researchers, and eventually from regulators — for standardized reporting on how much frontier model development is machine-led, in the same way the industry was pushed into reporting safety evaluations.

Kavukcuoglu himself has offered a practical yardstick that cuts through the benchmark noise: count the agents in your day. For DeepMind, the answer six months ago was zero agents in the training loop, and today is a supervised but autonomous slice of the experimental cycle. The next checkpoint is not whether agents can do more — the trajectory on that is not seriously in doubt — but whether the supervision layer, the control roadmap and the human review capacity scale at the same rate as the autonomy they are watching. That race, between capability and oversight inside the labs themselves, is now running in production.