Analysis
Demis Hassabis: Today's AI Agents Are a 'Practice Run' for AGI
The Google DeepMind CEO says artificial general intelligence could arrive by 2029 and warns the world is not prepared for how quickly these systems are advancing.
By Michael C ·
Demis Hassabis, the CEO and co-founder of Google DeepMind, told Axios on May 26, 2026 that today's AI agents — systems that can plan tasks, use software, and act across applications with limited human supervision — should be understood as a "practice run" for artificial general intelligence. The statement represents a significant sharpening of Hassabis's public position on AGI timelines. He has previously indicated he expects AGI around 2030. In the Axios interview, he said 2029 is now a possibility.
What Hassabis Actually Said
The core of Hassabis's argument is that the capabilities being demonstrated by current AI agents — the ability to decompose complex tasks, maintain context over extended interactions, use tools, and recover from errors — are not merely impressive product features. They are, in his framing, early demonstrations of the cognitive architecture that will eventually constitute AGI. The agents are not AGI. But they are practicing the skills that AGI will require.
Hassabis also warned that the pace of advancement is outrunning the world's ability to prepare for its consequences. "We're not prepared for how quickly these systems are advancing," he told Axios. This is not a new concern — Hassabis has been making versions of this argument for years. What is new is the context: he is now making it as the CEO of a company that has just demonstrated, at Google I/O 2026, a suite of agentic products that are already being deployed to hundreds of millions of users.
The Gemini Spark Context
Hassabis's comments came in the wake of Google I/O 2026, where the company unveiled Gemini Spark — an always-on AI agent that runs continuously in the background, monitoring information sources, sending emails, making purchases, and completing tasks on behalf of users. Gemini Spark is designed to spend your money and send your emails — a description that captures both the capability and the concern.
Google also announced that Gemini is being integrated directly into Search, replacing the traditional ten-blue-links interface with an AI-first experience that generates answers, books appointments, and completes transactions without requiring users to visit external websites. The implications for the broader web ecosystem — publishers, advertisers, e-commerce companies — are significant and largely unresolved.
The next wave of AI agents should be viewed as a dress rehearsal for AGI. We are not prepared for how quickly these systems are advancing.
Demis Hassabis, Google DeepMind CEO
The 2029 Timeline
The shift from "around 2030" to "possibly 2029" may seem like a minor adjustment, but in the context of AGI forecasting, it is significant. Hassabis has historically been one of the more conservative voices on AGI timelines among major AI lab leaders. His willingness to move the date forward reflects genuine acceleration in the capabilities he is observing internally at DeepMind.
For context: in 2022, most serious AI researchers would have placed AGI at 2040 or later. In 2024, following the release of GPT-4 and the rapid scaling of large language models, the median forecast among AI researchers had moved to the mid-2030s. In 2026, with agentic systems demonstrating capabilities that were considered science fiction two years ago, the most credible forecasters are now clustering around 2028–2032. Hassabis's 2029 estimate puts him in the middle of that range.
The Preparation Gap
The more urgent concern in Hassabis's remarks is not the timeline but the preparation gap. Governments, educational institutions, legal systems, and social safety nets were designed for a world in which human cognitive labor was the primary input to economic production. The transition to a world in which AI systems can perform most cognitive tasks — and are beginning to perform some physical ones — requires adaptations that none of these institutions have yet made.
Hassabis's own company is part of this dynamic. Google has deployed AI systems that are already displacing some categories of knowledge work. The company's AI Overviews feature in Search has reduced traffic to news publishers and reference websites. Gemini Spark will further reduce the need for human intermediaries in tasks like travel booking, customer service, and information research. DeepMind's AlphaFold has transformed structural biology in ways that are still being absorbed by the research community.
A Warning From Inside the Machine
What makes Hassabis's warning notable is its source. He is not an outside critic or an academic researcher. He is the CEO of one of the world's most capable AI research organizations, with direct visibility into the capabilities of systems that have not yet been publicly released. When he says the world is not prepared, he is speaking from a position of unusual epistemic authority.
The question his remarks raise — but do not answer — is what preparation would actually look like. Hassabis has previously argued for international AI governance structures, for investment in AI safety research, and for ensuring that the benefits of AI are broadly distributed rather than concentrated in a small number of companies and countries. These are reasonable positions. They are also positions that have been articulated by many people for many years without producing the institutional changes that would constitute genuine preparation. The gap between the pace of AI development and the pace of governance adaptation may be the defining challenge of the next decade.