Ask what finished, not how busy the AI looks
A long-running AI task may look active even when nothing is getting done. Mittal and colleagues study users’ reports of delegation failures, including stalled progress. Their findings are not a rate of failure across workplaces.
What the source reports
Mittal and colleagues study what can go wrong when people hand work to AI tools. The Beyond Prompting digest describes a map drawn from 163,487 Reddit posts and a study with 73 users. One problem is activity that continues without progress. Users also perceived more severe failures as harder to notice. The evidence does not establish how often these problems occur across all workplaces. The digest suggests asking for the last completed step and a time for the next check on a draft-only task. That is a proposed way to watch progress, not a proven cure. If no step has advanced, a person can pause the run before giving it more freedom.
Original source
- Title
- Babysitting AI Agents: A User-Centered Taxonomy and Study of Agentic Delegation Failures
- Author
- Shravika Mittal, Sunnie S. Y. Kim, Scott Counts, Jennifer Neville, Debarati Das, Philippe Laban
- Publication
- arXiv
- Date
- Monday, October 26, 2026