Creating Tools That Use Tools with Vibhaa Sivaraman

While we gear up for Season 2, we're sharing a special episode to bridge the seasons. Recorded live at SXSW, this conversation examines how AI agents are moving out of controlled research environments and into real-world consumer applications—and what it takes to make them reliable enough to matter. Cognitive scientist Danielle Perszyk is joined by Amanda Doerr, VP of Core Shopping at Amazon, Michael Giannangeli, Head of Agentic AI for Amazon Nova, and Michael Reiczyk, VP of Technology at Bandsintown, to discuss the gap between agentic capability and customer trust, including where users are willing to delegate decisions and where they pull back. The conversation covers reinforcement learning as a tool for improving model reliability, the shift from web-actuated to API-driven agentic shopping, and how human-in-the-loop design is shaping deployment across retail, live events, and foundation model development. Across all three domains, the panel finds that durable customer problems remain constant—even as the technical approaches to solving them change rapidly.

What would it mean to design AI for human flourishing? In the final episode of this season of “Making a Mind,” Cognitive scientist Dr. Danielle Perszyk sits down with Dr. Pat Pataranutaporn, Assistant Professor at the MIT Media Lab and founder of the Cyborg Psychology group, to explore how we move beyond optimizing models—and toward optimizing human development.  They discuss intelligence augmentation (IA) versus artificial intelligence (AI), why benchmarking model capability isn’t enough, and how we might instead measure AI by its impact on curiosity, learning, and collective well-being. From interdisciplinary meta-science to the risks of dehumanizing people while humanizing machines, they examine how AI can help us get smarter at getting smarter—without undermining what makes us human. 

Human minds are scaffolded by the tools we create, but what happens when we build tools that can use other tools?  Cognitive scientist Danielle Perszyk sits down with AI researcher Vibhaa Sivaraman to discuss agent tool use and the future of the digital world. They unpack what it really means to build computer-use agents—not just chatbots with function calls—by exploring personalization, multi-agent collaboration, and the idea of agents as the next “cognitive technology.” As agents begin navigating the web on our behalf, they examine how digital environments might evolve and whether agents should think like us, or complement us in entirely new ways. 

Pre-trained language models already contain vast knowledge—the challenge is producing the reasoning needed to handle ambiguous, multi-step tasks. Cognitive scientist Dr. Danielle Perszyk sits down with Amazon AI researcher, Meiqi Sun, to explore the shift from simple action execution to high-reasoning agents. Drawing parallels to human cognitive development, they discuss how reinforcement learning enables models to generate and refine their own chains of thought rather than relying on rigid, human-written templates. Together, they unpack why teaching agents to reason requires the freedom to explore, struggle, and self-correct. 

What if the key to building intelligent agents isn't just better models, but better teachers? Cognitive scientist Dr. Danielle Perszyk sits down with AI researcher Anirudh (Ani) Chakravarthy from Amazon's AGI Lab to explore how agents learn—not through memorization of data sets, but through structured experience.  Drawing parallels to human development, Ani introduces a training approach where two AI agents work together: one explores the web to discover tasks at the frontier of its capabilities, while the other learns from these challenges—a new approach to self-play. Together, Ani and Danielle discuss how this process points to a form of embodied intelligence distinct from language models—and what it could mean for the future of human-AI collaboration. 

A system that succeeds once is a demo. A system that succeeds every time is a breakthrough. Dr. Danielle Perszyk sits down with AI researcher Deniz Birlikci from Amazon's AGI Lab to explore how reinforcement learning (RL) is transforming AI agents from impressive demos into dependable tools that work consistently in real-world environments.  Danielle and Deniz discuss why reliability, not accuracy, is the true bottleneck for web agents, the critical role of a robust verification system, failure models that RL attempts to fix, and the extraordinary complexity of orchestrating live browsers with perception and actuation stacks. Discover how RL is building the foundation for agents that can handle complex workflows reliably alongside humans. 

Before an AI agent can reason or plan, it has to see. Dr. Danielle Perszyk and AI researcher Matthew Elkherj explore why user interface (UI) understanding is one of the most underestimated challenges in building autonomous agents—and why it’s foundational to creating reliable AI teammates. Danielle and Matthew discuss the distinct reliability requirements of agents, how perceptual hallucinations can be a feature (rather than a bug), and the role of synthetic gym environments in training. Together, they explain why building reliable agents requires solving interconnected challenges—from how agents perceive digital interfaces to how they learn from mistakes, handle real-world complexity, and ultimately augment human capacity. Please note: this podcast was recorded in August 2025.  

How did we get from language models to AI agents that can take action? Dr. Danielle Perszyk sits down with Kelsey Szot, product lead at Amazon's AGI Lab and one of the original founders of Adept (a startup that helped pioneer modern AI agents) to discuss the technical breakthroughs that transformed AI from pattern recognition to agentic capabilities. Danielle and Kelsey trace the evolution from early distributed training at scale to today's autonomous systems that can reason, plan, and interact with real environments—exploring the shift from rigid, rules-based automation to AI that can generalize across changing interfaces and complex workflows.

How do you build environments complex enough to train agents that can handle the real web? Dr. Danielle Perszyk sits down with Jason Laster, an engineer leading Amazon's AGI Lab's effort to build reinforcement learning (RL) gyms— simulated web environments where agents learn—to explore how environment development is as critical as models, data, and compute. The browser is one of the most complex worlds we could possibly train in, and this conversation unpacks why high-fidelity simulations that capture every UI quirk matter more than building thousands of basic environments. Discover how RL gyms are finally becoming practical at scale, why observability and verifiable rewards are essential for rigorous training, and why simulated environments beat the real web for developing reliable autonomous systems. 

Coming soon!