Div Garg

Div Garg, who publishes research as Divyansh Garg and uses the GitHub handle Div-Infinity, works on AI systems that learn and take actions. He co-founded and led MultiOn before founding AGI, Inc., where a July 2026 interview identified him as CEO and co-founder. His work spans imitation learning, reinforcement learning, and agents that operate applications through their screens.[3][4][6][9]

Garg dates a change in his approach to 2023–2024, while working on web agents. In his July 2026 interview with Unite.AI, he recalled that agents could perform browser actions when websites had a suitable structure, but stopped working when modals or animations departed from the expected DOM. Smartphone applications suggested a different approach: operate the screens as a person would, rather than depend on separate integrations.[9]

He describes phone control as three connected problems: a vision-language model must recognize interface elements; the agent must select actions despite ambiguity, partially loaded pages, dialogs, and errors; and taps must land accurately. Longer workflows add dependencies between steps, which he says made end-to-end model training essential. He described AGI, Inc.’s implementation as hybrid in July 2026: actions and light reasoning happen on the phone, while heavy reasoning happens in the cloud, with fuller on-device execution as the direction of development.[9]

In his January 2026 essay “Apps Are Dead”, Garg argued for consumer agents that make manual interfaces optional. He recalled watching his grandfather spend twenty minutes trying to book a doctor’s appointment online and asking why the user had to adapt to the interface. His July formulation clarified the ambition: “The apps never go away. The dynamic simply shifts.” Apps would remain capabilities that an agent composes around a user’s intent.[10][9]

In February 2025, Garg said he had recently spun out AGI, Inc. after founding and running MultiOn, with some initial MultiOn founding members involved. He contrasted MultiOn’s developer APIs and enterprise work with the new lab’s research on agent fine-tuning, reinforcement learning, and human–AI collaboration. A central priority was making agent applications reliable enough for consistent use.[8]

Garg described REAL Bench as a miniature Internet made from replicas of popular websites, with task-success scores and a leaderboard. His explanation emphasized task diversity and explicit evaluation harnesses and reward functions: a request such as booking a ride needs a way to determine whether the agent actually completed it correctly. He argued that evaluations are necessary to trust agents because they are “pretty much stochastic black boxes.”[8]

His trust argument also includes authorization and defenses against prompt injection. In 2025, he reported seeing an attack of this kind at MultiOn and proposed auditing, user authorization, and potentially two-factor authentication as guardrails. In July 2026, he described user approval for purchases, messages, and account changes: an agent could reach checkout but should wait before completing the purchase. He summarized the design principles as “human in the loop for anything substantial” and “reversibility wherever possible.”[8][9]

Before the AGI, Inc. spin-out, Garg explained MultiOn’s browser-agent strategy in Latent Space’s November 2023 OpenAI DevDay discussion. “I want to do things beyond APIs,” he said, describing interactions with personal accounts through existing logins and cookies. He planned to make the agent embeddable in other applications through an API, including a MultiOn GPT experience, while taking responsibility for identity, payments, and consequential browser actions.[1]

At that time, Garg’s assessment of GPT-4 Vision was restrained. He said MultiOn had early access to the Vision API for more than a month and had tried it on multiple websites, and that its parsing could compress most websites to fewer than 3,000 tokens. For its then-current browser workflow, adding vision was not “really game changing for us.” That judgment concerned MultiOn’s existing capabilities in 2023; his later screen-control account addressed failures around changing interfaces and the demands of smartphone applications.[1][9]

Garg’s earlier research supplies context for his work on learning and decision-making. His personal career history dates Cornell computer-vision research under Kilian Q. Weinberger and Bharath Hariharan to August 2018–May 2020, focused on camera-only depth estimation and 3D object detection. He also worked as a research intern in Apple’s Special Projects Group under Ian Goodfellow from March–September 2020, researching reinforcement learning, inverse reinforcement learning, and generative modeling.[3]

Garg coauthored IQ-Learn: Inverse soft-Q Learning for Imitation, published at NeurIPS 2021, with Shuvam Chakraborty, Chris Cundy, Jiaming Song, and Stefano Ermon. The method incorporates environment dynamics into imitation learning while avoiding adversarial training: a single Q-function implicitly represents both reward and policy. The authors reported improvements in required environment interactions and scalability in high-dimensional spaces, often exceeding threefold, in their evaluated offline and online imitation-learning settings.[6]

In Extreme Q-Learning: MaxEnt RL without Entropy, an ICLR 2023 oral paper, Garg and Joey Hejna were equal contributors alongside Matthieu Geist and Stefano Ermon. The method uses Extreme Value Theory to estimate optimal soft-value functions directly, without sampling from a policy. This avoids computing Q-values for out-of-distribution actions, a source of error in reinforcement learning. The authors reported gains of more than ten points over prior work on challenging D4RL Franka Kitchen tasks and moderate improvements over SAC and TD3 on online DM Control tasks.[7][3]

Garg also taught Stanford’s Fall 2021 CS25: Transformers United alongside Chetanya Rastogi and Advay Pal, with Chris Manning as faculty advisor. The seminar combined instructor lectures, guest talks, and discussion to examine transformer architectures across language, vision, reinforcement learning, and biology, with the stated aim of encouraging research across those communities.[5]