Rich Sutton's Oak Lab: AI agents that learn on their own

The timeline: from Montreal to the Toronto garage

Richard Sutton is not a new name in the AI world. He is the co-author of the go-to book on reinforcement learning, written with Andrew Barto. In 2024 he received the Turing Award , the most prestigious recognition in computer science. However, instead of retiring to academia, he chose to found a new startup: Oak Lab , based in Toronto.

The news was reported by The Decoder on July 13, 2026. Sutton's move comes at a precise moment. In fact, the AI agent market is already crowded with solutions based on static LLMs. On the contrary, Oak Lab focuses on something structurally different: agents that continuously learn from interaction with the environment.

This approach brings back the roots of classical reinforcement learning. At the same time, it launches it into a modern infrastructure setup. So, it's not just a simple tweak to existing models.

The technical catch: why deep learning "is not enough"

Sutton has defined current deep learning methods as "weak and inefficient." It's a strong statement. Therefore, it's worth understanding what he means in detail.

Large language models — GPT, Claude, Gemini — learn during the training phase. After deployment, they remain static. They don't update their knowledge based on subsequent interactions, except for periodic fine-tuning mechanisms. Consequently, they require frequent human intervention to stay up-to-date and relevant.

A continuous reinforcement learning-based agent, on the other hand, adapts in real-time. It learns from environmental feedback. Therefore, over time, it improves its performance without the need for new centralized training cycles. This feature is the substantial difference that Oak Lab intends to leverage. To further explore the technical distinction between static LLMs and continuous learning systems, the MIT Technology Review offers an up-to-date picture of the state of research.

Winners and losers in the new autonomous agent ecosystem

The founding of Oak Lab reshapes the competitive balance in the industry. However, not everyone comes out of it strengthened in the same way.

Who gains ground: companies that are building flexible infrastructures capable of integrating next-generation agents. Furthermore, automation platform vendors who can update their APIs to communicate with continuous learning agents will have a structural advantage.

Who risks falling behind: AI marketing solutions based on rigid workflows and non-updatable pre-trained models. In fact, if the industry moves towards autonomous agents, closed architectures become a limitation. Similarly, marketing teams that today depend on a single vendor risk finding themselves tied to already outdated technologies.

According to the analyses of Gartner , by 2027-2028 autonomous AI agents will be integrated into major enterprise platforms. The direction indicated by Sutton accelerates this trajectory.

The SHM Studio take: what changes for operational marketing

In SHM Studio we track the evolution of AI agents with a special eye on what it means day-to-day for Italian marketing teams. So, we want to give you a hands-on read, not just textbook theory.

The first impact concerns content personalization . An agent that learns continuously can adapt messages, offers, and formats based on actual user behavior. Not on aggregated historical data, but on real-time signals. This changes the logic behind digital marketing strategy and production of SEO content .

The second impact concerns campaign management . Today, optimizations on Google Ads and LinkedIn Ads happen on weekly or monthly cycles. With autonomous agents, optimization becomes continuous and adaptive. Consequently, the role of the marketing manager shifts: less manual execution, more strategic supervision.

The third impact concerns customer intelligence . Agents capable of interacting and learning can build much more accurate predictive models of customer behavior. This strengthens activities in AI applied to marketing that we already offer to our clients today.

The work still in progress: limits and unknowns of Oak Lab

Still, we've gotta keep a realistic mindset. Oak Lab is a brand-new startup. They don't have commercial products out yet, nor any real-world success stories at a big scale.

Continuous reinforcement learning presents non-trivial technical challenges. First and foremost, learning stability in real-world environments is difficult to guarantee. Furthermore, data privacy management — fundamental in European contexts subject to GDPR — becomes more complex when the agent learns incrementally from user interactions. Therefore, Italian companies will need to carefully evaluate the regulatory implications before adopting these technologies.

Subsequently, we'll need to understand how Oak Lab intends to position itself against the big players: OpenAI, Google DeepMind, Anthropic. All are working on agentic systems. Sutton's difference is the theoretical approach. But the distance between research and enterprise product is often longer than expected.

Next moves: how to get ready today

For marketing and digital managers of Italian companies, the practical question is: what to do now, before these agents arrive on the market in commercial form?

First of all, it is worth map repetitive marketing processes which today require frequent human intervention. These are the natural candidates for agentic automation. For example, managing responses on social channels, dynamically updating landing pages, or continuous lead segmentation.

Second, it is useful build flexible data infrastructures . An autonomous agent only works well if it has access to clean, organized, and live data. Because of that, putting money into solid data setups today—hooked up with the website and with CRM systems — is a choice that will pay off even in the short term.

Third, it is strategic monitor how existing platforms evolve . Google, Meta, and LinkedIn will integrate agentic capabilities into their advertising platforms. Those who already understand the mechanics of automatic optimization will have an advantage in adopting the more advanced versions. Our activities of SEO and Digital marketing already factor in this trajectory.

Finally, it is advisable train internal teams on the difference between generative AI and agentive AI. It's not about technical details. It's about understanding which type of tool is suitable for which type of problem. For this, the SHM Studio blog will continue to publish operational analyses on these topics.

For those who want to dive deeper into AI agents from a strategic angle, the report by McKinsey on the state of AI offers up-to-date data on the pace of enterprise adoption globally. Anyone wishing to connect with the team at SHM Studio on these topics can do so through our contact channels.