Most AI agents wait for you to ask for help. This means you must recognize when you need assistance and know how to request it, which leads to many missed opportunities. A proactive agent that offers help without being asked could be more useful, but it must learn when you actually want assistance and when you would find it disruptive.
The researchers built a desktop agent called Coco that constantly watches what you do on your computer. A separate Observer Agent processes this activity, builds a compact model of your current work, and predicts when help would be welcome. It learns by noting whether you accept or dismiss its suggestions and by observing your work patterns, all without sending any of your activity data off your device.
In a week-long study with ten participants, Coco's ability to correctly time its help improved from a score of 0.199 to 0.515. This shows that an agent can learn each person's preferences for when to offer assistance, making proactive help more useful and less annoying.