When you give an AI coworker a 'defined role,' who actually decides what it works on each day — you, the company, or the AI itself?

This explores who sets the daily agenda for a persistent 'coworker' AI agent: the person working alongside it, the organization that defines its role, or the agent itself. The short answer is that the corpus describes the pieces without a clear owner. Google's Gemini announcement splits agents into two kinds. Temporary sub-agents disappear after one task. Coworker agents have a company email address, their own storage and a 'defined role', and they see only the context someone explicitly shares with them How does Google separate temporary agents from persistent ones?. A role can be defined and context can be shared, but the announcement never says who turns that role into a daily to-do list.

The most useful lens comes from studies of human-AI teams. In a study of 244 BCG consultants, the way people split work with AI didn't depend on the tool or the task. It depended on two questions: who picks the task, and who picks the method Do humans and AI divide labor differently in knowledge work?. Applied to coworker agents, 'who decides the daily work' is really two separate decisions, what to do and how to do it, and they can be handed out differently. Research on 'autonomy alignment' makes the same point from the design side. How much decision-making power an agent has should match what users expect. That power is set in three places: the model (when it escalates to a human), the harness (the software that enforces the rules) and the interface (how delegation is shown to users) How should AI agents balance autonomy with user oversight?.

The surprise is that assigning work may matter less than you'd think, and supervising it may matter more. In a randomized trial of Copilot across 66 firms, users saved about two hours a week on email, but the kind and amount of work they did stayed the same. The authors' explanation is that most work depends on team coordination that hadn't changed yet Does Copilot change what knowledge workers actually do?. An analysis of more than 73,000 Reddit posts about the OpenClaw agent found a similar pattern. Users were mostly satisfied with what the agent delivered, but their values broke down during supervision, in the gap between setting the agent up and reviewing its results Where do user values break down in agent supervision?. Nielsen argues that interfaces need a separate layer for overseeing and steering work, apart from the layer where you state what you want How should AI interfaces handle the shift from doing to supervising?.

When several people share agents, the question gets harder. In tests across five frontier models, one coordinator agent serving everyone produced better group outcomes than a team of personal agents, one per user. The per-user teams sometimes fell apart entirely Do teams of personal agents outperform a single coordinator?. That points toward central assignment, but central assignment brings governance gaps of its own. Whether agents passing work to each other counts as a violation is defined only by written policy, and when there is no policy, every handoff is allowed by default How do policies determine whether agent transfers are violations?. When delegation chains fail, the harm can land on people who never gave the agent an instruction Who actually bears the risk when multi-agent workflows fail?.

The corpus has no study of how coworker agents actually get their daily work in real deployments, so this remains an open question. What it does give you is a way to think about it. 'Who decides' splits into choosing tasks, choosing methods, setting escalation rules and reviewing results, and the hard design problems are in the last two.