Key Takeaways
- Early AI agent work, like Nova Act, focused on automating "atomic interactions"—think precise clicks and scrolls—to achieve reliability.
- Danielle Perszyk argues that true AI agent reliability isn't about perfectly executing tasks, but about modeling the user's mind: understanding their dynamic preferences and intentions.
- User goals aren't static; interaction with software actively reshapes and refines what a user is trying to achieve, making fixed task execution insufficient.
- This means building AI more like a human executive assistant—one who grasps the underlying why of a request—rather than a rigid robotic process automation tool.
From 'Click Reliability' to 'Mind Modeling': The New AI Standard
For a long time, the pursuit of reliable AI agents felt like an exercise in digital puppetry. Researchers, including Danielle Perszyk in earlier work like Nova Act, focused on perfecting what she calls "atomic interactions." This meant an AI agent reliably clicking, scrolling, and inputting data in exactly the right place, every single time. The goal was to meet “the models where they were at”—to get the machine to perform simple, repeatable actions with zero error. It sounded like the definition of reliable automation.
But that definition, Perszyk explains, falls apart the moment you consider any real-world human task. Take something as common as booking travel online. Your goal isn't just to click buttons in a certain order; it’s to find the best flight for you, balancing cost, convenience, and a dozen other preferences that might shift during the booking process itself. “Everything that we do,” Perszyk says, “unless you're using like the arbitrary software skills for submitting your invoices for for work... we are actively thinking about and the interaction with computer is shaping and refining the way that we are thinking about the the goal itself.”
This is the critical pivot: the user's intent is not a fixed input. It's a living, breathing thing that evolves as they interact with the software. An agent that simply executes a pre-programmed sequence, no matter how flawlessly, isn't reliable if it can't adapt to that evolving intent. It's like having a perfectly efficient chef who only knows how to make one dish, even if you change your mind about what you want for dinner.
Building Agents That Actually Understand You
For Perszyk, the difference between a truly reliable agent and a glorified macro comes down to one thing: the agent's ability to model the user's mind. “The person understands the human users mind and and their goals,” she states, “and they can decompose not just the task but the the preferences and the intentions of the human.” This elevates the agent from a simple tool to something more akin to a human executive assistant. An assistant doesn't just follow instructions; they anticipate needs, clarify ambiguities, and adapt to your changing priorities because they understand your broader objectives.
This shift in perspective completely reframes how founders and builders should approach AI. Reliability is no longer measured by how accurately an agent can click a button, but by how well it grasps and responds to the often unstated, always evolving, preferences and goals of its user. “Ultimately reliability has less to do with clicking in the same place and scrolling and more to do with modeling the user's mind,” Perszyk concludes. “And that shift is everything. That reframes how we think about what it is that we're building.”
Ignoring this means building brittle agents that fail at the first sign of a user's evolving thought process. Embrace it, and you start designing systems that feel genuinely intelligent, perceptive, and, yes, truly reliable.
What to Do With This
Founders building AI agents: Immediately audit your agent's evaluation metrics. If you're primarily scoring on perfect execution of atomic tasks (e.g., successful clicks, form fills), you're measuring the wrong thing. Instead, design your agent's training and feedback loops to prioritize goal clarification, preference learning, and dynamic adaptation to user intent shifts. Start instrumenting your agent to detect when a user's interaction patterns deviate from initial task definitions, signaling a goal evolution your agent must learn to accommodate, not just override.