AI agents that act before you ask could save you effort — or just hand you a new job: watching them.

This explores whether agents that take initiative can lift the overhead around real work: the clarifying, coordinating, and keeping track. The corpus has no study aimed squarely at 'work about work' as a category. It does have a fairly consistent picture of what happens when agents get proactive, and the answer is a qualified yes. Proactivity removes effort, but it also creates a new kind of overhead: watching what the agent is doing.

The upside is real. In simulations, agents that offer relevant information before being asked cut the number of conversation turns by up to 60 percent. That is the back-and-forth humans normally skip by anticipating each other's needs Could proactive dialogue make conversations dramatically more efficient?. AI systems rarely do this, and the reason is surprising. It isn't that they can't. Training that rewards a good next reply quietly removes initiative. Proactive behaviors like asking clarifying questions can be trained back in, jumping from under 1% to nearly 74% with reinforcement learning Why do AI agents fail to take initiative?. So the passivity that forces users to spell out every step is a design choice, not a fixed limit.

The catch shows up in the field. In a study of 18 programmers, proactive coding agents reduced user effort compared with prompt-only tools, but they also broke into people's workflow Do proactive AI coding agents actually disrupt programmer workflow?. The fix wasn't less proactivity. It was visible presence cues and shared context, so people could see what the agent was up to without having to check on it. That lines up with the argument that intelligence and adaptivity alone produce 'socially blind' agents, and that civility (good timing, respecting boundaries, not overriding the user) is what makes initiative welcome rather than one more thing to manage How can proactive agents avoid feeling intrusive to users?. One proposed route to good timing is reading cognitive state from behavior like hesitation, gaze, and typing speed. The same signals that let an agent wait for a good moment could also be used to profile and manipulate users Can AI systems read cognitive state from interaction patterns alone?.

Another lever is making the split of work explicit instead of automatic. Cocoa lets users assign each plan step to themselves or to the agent, and revise the plan mid-task. Users got more control without extra effort, and field users especially valued spelling out who does what Can users steer AI agents better by revising plans mid-execution?. One theoretical piece flips the question: agents that can make decisions may start handing tasks back to humans Can AI systems delegate tasks back to humans?. If that happens, an agent could just as easily generate work about work as absorb it.

The point you might not expect: lower cognitive load isn't always the goal. When people stop thinking, they over-rely on the AI. Designs that deliberately add friction cut that over-reliance, but users rated them least favorably, and the human-AI teams still did worse than the AI alone Do forcing functions reduce AI overreliance without harming team performance?. The best proactive agents probably won't remove all effort. They'll take over the busywork and leave the judgment calls with you.