From OpenAI to Nvidia, researchers agree: AI agents have a long way to go

Only in the Bay Area does spending a Saturday geeking out about AI agents—alongside 2,000 students, researchers, and tech insiders crammed into UC Berkeley—feel like a totally normal weekend plan. As I picked up my badge at the day-long Agentic AI Summit and watched the line snake through the student union lobby, it felt less like an academic conference and more like Silicon Valley’s version of a buzzy New York brunch spot.

This was certainly due to the speaker lineup, which was stacked with top AI researchers and scientists, including Jakob Pachocki, chief scientist at OpenAI; Ed Chi, VP of research at Google DeepMind; Bill Dally, chief scientist at Nvidia; Ion Stoica, cofounder at Databricks & Anyscale, as well as a UC Berkeley professor; and Dawn Song, a pioneering UC Berkeley professor focused on AI security. 

The popularity also might have been due to the buzzy topic—AI agents, generally defined as an AI-powered system that can complete tasks, mostly autonomously, using other software tools. Think not only suggested a vacation itinerary, but also booking the flight and making the hotel reservation.

As my colleague Jeremy Kahn said in a recent article, “This kind of automation is a perennial C-suite fever dream. Over the past decade, companies embraced ‘robotic process automation,’ or RPA. This was software that could automate repetitive tasks, such as cutting and pasting between database programs. But traditional RPA systems are inflexible and unable to deal with exceptions, and can usually handle only one narrow task.” Agentic AI is meant to be both more flexible and powerful, adapting to business needs.

In a January 2025 blog post, OpenAI CEO Sam Altman said, “We believe that, in 2025, we may see the first AI agents ‘join the workforce’ and materially change the output of companies.”

But despite the hype, the overall message at the Agentic AI Summit was cautious and grounded: Agents may be the buzziest trend in AI right now, but the tech still has a long way to go, they said. AI agents, unfortunately, aren’t always reliable. They may not remember what came before.

Google DeepMind’s Chi, for example, stressed the gap between what agents can do in curated demos versus what’s still needed in real-world production environments. Pachocki highlighted concerns around the safety, security, and trustworthiness of agentic systems, particularly when they’re integrated into sensitive applications or operate autonomously. 

“I still don’t think agents have really lived up to their promise,” said Sherwin Wu, head of engineering at OpenAI API. “Certain more generic cases have worked, but my day-to-day work doesn’t really feel that different with agents.”

While today’s agents may not currently live up to the massive hype (consider Salesforce CEO Marc Benioff’s recent claim that a shift to digital labor means he will be the “last CEO of Salesforce who only managed humans”), the speakers at the Agentic AI Summit still had plenty of optimism to share. Databricks’ Stoica expressed enthusiasm about infrastructure improvements that are making it easier to build agentic systems. Nvidia’s Dally suggested that continued hardware advances will enable more powerful and efficient agent behavior. Several pointed out “narrow wins” in specific domains, like coding.

Today’s AI agents may still have growing pains, but given the crowded UC Berkeley ballroom, the industry maintains its eye on the prize: AI agents that can reliably operate in the real world. The payoff, they believe, will be well worth the wait.

With that, here’s more AI news.

Sharon Goldman
sharon.goldman@fortune.com
@sharongoldman

AI IN THE NEWS

U.S. agency approves OpenAI, Google, Anthropic for federal AI vendor list. Reuters reported today that the General Services Administration, which is the U.S. government’s central purchasing arm, added OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude to a list of approved AI vendors in order to accelerate use of the technology by government agencies. The tools will be available to the agencies through a platform with contract terms in place. The GSA said approved AI providers “are committed to responsible use and compliance with federal standards.”

The AI spending boom could have real consequences for the U.S. economy. According to the Washington Post, Big Tech’s record-breaking investment in artificial intelligence—more than $350 billion this year from Google, Meta, Amazon, and Microsoft—is becoming a major economic force, even as the broader U.S. economy shows signs of slowing. While job growth is cooling, this massive AI spending spree is fueling construction of data centers and driving demand for chips, servers, and networking gear—potentially boosting GDP growth by up to 0.7% in 2025. But economists warn the growing reliance on tech giants to prop up the economy is risky: if the AI boom loses steam, the economic fallout could be significant. 

AI sales tool Clay raises $100 million at a $3.1 billion valuation. The New York Times Dealbook reported that Clay, which helps sales reps and marketers find new leads and turn them into customers, has raised $100 million at a $3.1 billion valuation.The round was led by CapitalG, an investment arm of Alphabet, Google’s parent company. Other participants included Meritech Capital Partners and Sequoia Capital. It comes around six months after the start-up raised money at a $1.25 billion valuation.

EYE ON AI RESEARCH

Google DeepMind’s new Genie 3 ‘world model’ creates real-time interactive simulations. Google DeepMind has unveiled Genie 3, a powerful new AI system that can generate rich, interactive virtual worlds from simple text prompts—making it possible to navigate dynamic environments in real time at 24 frames per second. But while it’s tempting to immediately leap to using the model for the ultimate gaming experience, it’s actually the latest leap in the company’s long-term push toward ‘world models’—or AI systems that can learn how the world works and simulate real-world environments. These are seen as key to training advanced agents and, eventually, achieving artificial general intelligence. Unlike prior video generators, Genie 3 allows users to move through AI-generated environments that stay visually consistent over several minutes—and even respond to commands like “make it snow” or “add a character.” For now, DeepMind is limiting access to Genie 3 to a small group of researchers and creators while it explores responsible deployment and risk.

FORTUNE ON AI

North Korean IT worker infiltrations exploded 220% over the past 12 months, with gen AI weaponized at every stage of the hiring process —by Amanda Gerut

AI is doing job interviews now—but candidates say they’d rather risk staying unemployed than talk to another robot —by Emma Burleigh

These charts show how China is pulling ahead of the U.S. in the race to power the AI future —by Matt Heimer and Nick Rapp

AI CALENDAR

Sept. 8-10: Fortune Brainstorm Tech, Park City, Utah. Apply to attend here.

Oct. 6-10: World AI Week, Amsterdam

Oct. 21-22: TedAI San Francisco. Apply to attend here.

Dec. 2-7: NeurIPS, San Diego

Dec. 8-9: Fortune Brainstorm AI San Francisco. Apply to attend here.

BRAIN FOOD

Could “depth of thought” be key to AI reasoning? 

A tiny new AI model is challenging what we know about how models learn to reason: Researchers from Singapore’s Sapient Intelligence recently released the Hierarchical Reasoning Model (HRM), which draws inspiration from the brain’s layered thinking process—and the results have the AI community chattering. Despite being 100 times smaller than ChatGPT and trained on just 1,000 examples (with no internet data or step-by-step guidance), HRM solves tough logic problems like Sudoku, maze navigation, and abstract reasoning tasks that stump much larger models. Instead of mimicking human language, HRM reasons internally—quietly working through problems in hidden loops, much like a person thinking through a puzzle in their head. Its success hints at a radical shift in AI: one where depth of thought might matter more than scale.