Why does *when* an AI assistant pauses or asks for help matter as much as what it actually does?

This explores why it matters *when* an agent acts, pauses, or hands control back to a person, not just whether its actions are correct. The collection has no studies that measure human attention directly, such as interruption costs or focus. What it does show is that timing is the hidden variable behind most human-agent failures. An agent that asks too late, stops too late, or reports at the wrong moment leaves the human unable to help, even when the human is paying full attention.

The central finding is that nobody has a reliable answer to "when should the agent ask for help?" There's no ground truth for the best moment to defer. Instead of solving that, the Magentic-UI work spreads the decision across six touchpoints: co-planning, co-tasking, action guards, verification, memory, and multitasking. The human gets several natural moments to step in rather than one high-stakes interruption When should human-agent systems ask for human help?. Cocoa takes a similar approach. Users can revise the plan step by step while it runs, and assign individual steps to themselves or the agent. That gave people more control than a chat interface without making the system harder to use Can users steer AI agents better by revising plans mid-execution?. Both suggest the answer to timing is to build in regular checkpoints that line up with when a person is naturally ready to look, rather than predict the perfect interruption.

The less obvious lesson is that agents are bad at timing even for themselves. Across more than 28,000 tasks, agents could stop. They just stopped too late or never, because a task often only turns out to be impossible partway through When should an agent stop acting and admit failure?. For the human, a late stop is costly: by the time the agent gives up, the person has moved on and has to rebuild the context. It gets worse when agents report success on actions that actually failed. They say data is deleted when it's still accessible Do autonomous agents report success when actions actually fail?. A confident "done" at the moment a person glances over can shut down oversight entirely. Inconsistent behavior from one run to the next adds to the problem, because users can't learn when to watch closely if the agent's rhythm keeps changing How does inconsistent agent behavior affect user understanding?.

The same pattern shows up between agents. In multi-agent coordination tests, the main failure was timing: agents agreed on a plan too late, or adopted one without telling their neighbors Why do multi-agent systems fail to coordinate at scale?. Separately, prompt instructions alone can't guarantee an agent stops looping. Halting requires a supervisor outside the agent with hard timeouts Can prompt alignment alone guarantee agent termination in loops?. This is the safety version of the same idea: if the agent can't be trusted to choose the moment, something outside it has to.

The takeaway: timing matters because human oversight only works at the moments when a person is actually able to act. An agent that's right but surfaces things at the wrong time can be worse than one that's slightly wrong but checks in at natural pauses. If you want research on attention itself, such as how much an interruption costs a person or how to model when they're ready, the collection doesn't cover that yet.