Faria Huq and the Reason Behind a Human Step

One figure in the CowPilot paper follows an AI agent, a program that operates a web browser on someone's behalf, through a small task: count the comments under the latest post in a space-themed forum. The agent finds the list of forums, scrolls twice, and never locates the right one. A person takes over, navigates to the forum, and hands control back. The agent then finds the post and counts.

The system logs that rescue as a "human step," and it would log a very different moment the same way. In the team's follow-up CowCorpus study, participants asked the agent to buy toothpaste without naming a brand. Only when they saw its pick did they realize the brand mattered to them. The forum step repaired a failure, though the record cannot say whether that failure belonged to the agent or to a forum whose own design may have been the obstacle. The toothpaste step was the person's task arriving late. A log entry is also where responsibility gets assigned, so filing both steps under one label obscures that they belong to different parties.

Faria Huq and her colleagues at Carnegie Mellon built the instrument that produces these entries. Zora Zhiruo Wang is her co-first author on the second paper, and Jeffrey Bigham and Graham Neubig advise. CowPilot is a browser add-on: the agent proposes each action, and a person can pause it and take that step instead. On 27 tasks from WebArena, a benchmark built from realistic mock websites, GPT-4o working alone succeeded 48% of the time. With a person able to step in, success rose to 95%. The people stepping in were three of the paper's own authors, so the result shows what the arrangement allows in expert hands.

CowCorpus widened the study to 20 participants and 400 recorded tasks. Working from notes participants wrote afterward, the team identified three reasons people stepped in: the agent made a mistake, its choice didn't match what the person wanted, or the page itself got in the way, with dropdown menus and CAPTCHAs that agents handle badly. The paper illustrates each reason but doesn't report how often each one occurred.

The team then trained a model to predict a single thing: will this person intervene at this step? All three reasons became one yes. The published methods attach no reason to individual interventions, and with 400 recordings that is a defensible simplification, though it gives up the reason. An agent taught to pause before a stubborn dropdown and one taught to pause before choosing a brand receive the same lesson, although the first needs better navigation and the second needs to learn something about a person.

The next system, PlowPilot, uses that prediction to decide when to check in. Sixteen people rated it 36.8% more useful than CowPilot. Those ratings record impressions, and only the four returning participants are clearly described as comparing against their own earlier CowPilot ratings.

Huq's September thesis proposal ranges more widely, but one line in its summary describes the moments the classifier folds into a single yes:

Read that way, some interventions are instructions, given in the only form available at the moment the person was able to give them.

That affects where a team puts its effort. A cluster of human steps on one website may be describing the website, while a cluster around one user may reflect a task still taking shape. When the logs file both as agent error, the engineering work goes to the model either way, and a team can spend a quarter teaching its agent one site's dropdowns while the preference it kept missing never appears in a report.