Danielle Perszyk

Danielle Perszyk is a cognitive scientist. In July 2025, she described leading a team at Amazon’s AGI Lab generating human-computer interaction data to train agents to collaborate with people; Amazon identified her as head of its HCI team in January 2026. Her public argument is that useful agents should strengthen human thinking and agency, with people retaining control over their work.[2][3]

In her July 2025 article, Perszyk criticized the attention and cognitive effort consumed by switching windows, searching message threads, and repeatedly entering information across systems. She framed the lab’s objective this way: “Instead of making AI smarter and giving it more agency, we’re building AI that makes us smarter and gives us more agency.” Her longer-term goal is for agents to infer people’s higher-level goals, which she argues requires modeling human minds and the social world alongside physical and digital environments.[2]

Perszyk explained Amazon Nova Act, an agentic model paired with an SDK, as the team’s initial approach to building toward that goal. Developers divide repetitive workflows into small computer interactions, such as clicking icons or typing into search fields, and combine probabilistic model calls with deterministic Python integrations. In her account, this structure addresses a practical limitation: models could perform small interactions more reliably than they could independently decompose and execute complex goals.[2]

Her example—“Click on the button for choosing the highest rated option”—shows how a model call can adapt to different websites while Python supplies the surrounding structure, including parallel execution across sites. She compared this combination to language: grammatical structure supports flexible meanings that context helps disambiguate. Developers may need to refine an instruction’s wording before composing it into a larger workflow.[2]

Perszyk reported in July 2025 that the team was creating thousands of simulated environments representing varied digital use cases to support reinforcement learning. She also proposed evaluating more than agent task performance: human creativity, collaboration, strategic thinking, abstract reasoning, learning efficiency, flow, and enjoyment of work should matter. These were development and measurement priorities, rather than reported demonstrations of improved human outcomes.[2]

Perszyk hosts Making a Mind, Amazon’s podcast launched on January 7, 2026, exploring the science needed to build useful general intelligence. Her conversations include Jason Laster on high-fidelity browser-training environments, Meiqi Sun on agent reasoning and self-correction, and Pat Pataranutaporn on evaluating AI through human curiosity, learning, and well-being. These discussions connect technical agent development with questions about how people benefit from using the systems.[3]

She also appeared on Latent Space in “Why AI Agents Don’t Actually Understand You”, published on July 11, 2026.[1]

Before her agent work, Perszyk studied the relationship between language and thought in infancy. Northwestern lists her among its 2017–2018 Ph.D. recipients, with the dissertation Linking Language and Cognition in Infancy: The Roles of Nature and Nurture, advised by Sandra Waxman. This earlier research supplies context for the language-and-thought specialization she later brought to human–agent collaboration.[4][2]