Raphael Shu

AI agents are becoming capable of doing substantial work, but turning individual capability into effective teamwork remains an open problem. The next challenge is not simply how one person uses one agent, but how teams of humans and agents work together. How should they share context, divide responsibilities, reuse expertise, make decisions, and remain accountable for the result? This talk presents the vision behind OpenAgents: a collaboration OS that gives human teams and agent teams a shared environment for getting work done. People can bring their own agents and private context, choose what to share, and collaborate through persistent threads, tasks, and artifacts. Agents can work over long periods, call specialists, hand off responsibilities, and escalate decisions to the right person. Humans set direction, contribute judgment, review results, and remain responsible for the outcome. We will examine concrete scenarios, including a software change moving through agent review, human approval, deployment, and testing, and a specialist agent whose expertise becomes reusable by colleagues and their agents. These examples reveal the infrastructure mixed teams need: shared context, clear ownership, reliable handoffs, permissions, and continuity across people, agents, and runtimes. The talk then considers how open collaboration interfaces could extend this model beyond a single application and support a broader ecosystem of human-agent teamwork.