Cling beats full-agent automation by letting humans steer each step

Researchers built Cling, a browser tool for spreadsheet tasks that span multiple web pages. Instead of handing the whole job to an AI, Cling breaks work into small steps and lets the human approve each one before the agent moves forward. A controlled study found that approach cut task time by more than half compared to doing the work by hand. It also beat fully automated AI that ran without human input. The study gives product teams hard numbers on a question that often stays theoretical: how much control is too much, and how much automation is too little? Cling sits in the middle. The human stays in the loop step by step, but the agent does the repetitive moving and copying. Neither party holds the whole job alone. For designers, the key finding is that small-step human direction outperformed both extremes tested. Full automation failed on complex tasks because errors compounded. Pure manual work was slow. Cling's approach, where the human points and the agent acts on that one step, kept errors low and speed high. That tradeoff is now documented in a controlled study, not argued from first principles.