Ben Stein

Ben Stein co-founded SuperDuper, which built personalized AI software for family logistics, and Teammates, which created AI coworkers for workplace collaboration. His career has taken him from medical computer vision and mobile messaging to developer platforms, climate technology, and autonomous agents.

Early in his career, Stein worked on machine-learning systems that helped radiologists identify potential lung-cancer nodules in CT scans. He subsequently co-founded Mobile Commons, an SMS-marketing company acquired by Upland Software, and spent seven years at Twilio in leadership roles spanning messaging, developer experience, enterprise products, and Twilio.org research and development. As chief product and technology officer at Arcadia, he worked on software and APIs for a cleaner, more decentralized electrical grid. He also co-founded QuitCarbon, focused on helping households transition away from fossil fuels.

In 2024, Stein and fellow former Twilio product leader Kenny Hoxworth introduced Teammates. Its virtual coworkers had their own identities, email accounts, and access to tools including Slack and Google Workspace. When a customer asked whether an AI coworker could respond to a Google Docs comment, Stein realized the answer depended on permissions, notifications, document access, and model behavior that nobody had explicitly specified. His AI Engineer World’s Fair talk developed that uncertainty into a practical approach to building autonomous products.

• Agent affordances over feature checklists. Specify capabilities such as communicating, collaborating, and handling documents, then discover how they combine. Unexpected behavior becomes something to investigate and constrain, not simply an omission from a feature specification.

• Evaluations as living product specifications. Use explicit behavioral criteria, acceptable pass rates, and model-assisted assessment to make judgment, tone, and other probabilistic qualities legible to product managers and engineers. Rapid prototypes reveal interaction problems, such as excessive clarifying questions, that conventional specifications miss.

• Software that supplies its own context. Stein argued against making users prompt engineers, favoring systems that learn through ordinary interaction. When Teammates customers struggled to provide enough context for agents to act, he shifted toward software that interprets existing information and proposes useful work proactively.

• Least-privilege agent architecture. After a virtual employee received a phishing email, Stein advocated narrowly scoped agents and restricted permissions. At SuperDuper, software handling private family information and untrusted email deliberately lacked outbound actions, limiting opportunities for prompt-injection-driven data exfiltration.

SuperDuper extracted schedules, deadlines, and action items from school notices and family email, creating a shared view intended to reduce the administrative burden concentrated on one caregiver. The company subsequently announced an agreement to be acquired by Life360 and scheduled its standalone application to close on August 17, 2026, without transferring customer accounts or data.