Automating the Lookup, Not the Inquiry
Cling’s prepared-task speed advantage supports automating repeated browser lookups and transfers while people keep directing the larger inquiry.
One user of Comet, an AI browser, described a benefit other than finishing faster. In a discussion of configuring Zoho One, a business-software suite, the commenter Powerful-Cheek-6677 reported handing over roughly 70 setup tasks or fields so they could work on other things. They also reported difficulties with dragging, clicking fields and closing dialogs. Another commenter reported poor results with Zoho setup. These are conflicting personal accounts, but the first makes the objective clear: getting away from the work.
The October 19 preprint presents Cling, a browser prototype for matching, filtering and transferring selected information. In repeated list-matching, distance-checking and paper-transfer tasks with 12 university students, Cling roughly halved reported duration against manual work and a Browser Use agent. Geometric averages, suited to proportional comparisons, counted timeouts as 240 seconds; comparators timed out more often. Local archived pages had their structure cached, excluding preparation cost.
A lookup can have a clear ending even when the inquiry does not.
Two jobs inside one row
Consider a literature-review spreadsheet. In a 2016 account of researching bottled water, researcher Raul Pacheco-Vega describes a sheet with citations, article summaries, personal notes, links between references, quotations and source pages. His notes include whether he agrees or disagrees with an idea. The sheet holds both records of what he read and his developing response to it.
Those columns can sit beside each other without being interchangeable jobs. Carrying a citation into the right row preserves an identifier. Recording a disagreement requires something the reader has made of the source. An assistant could usefully do the first without being asked to do the second. Automating the transfer would not, by itself, make the reading redundant.
Cling’s bundled comparison cannot isolate human control from preparation or specialization. The agent needed fewer clicks and tab switches; accuracy differences among completed attempts were not statistically significant.
A useful question can change
A 2020 field deployment of Mesh, a product-comparison tool, followed 5 people working on planned purchases over 1–2 weeks. All had used spreadsheets for online research. Their activity logs recorded additions and removals of products and comparison criteria.
One participant researching printers found little difference between the options on ink. They turned toward what they called “air printing,” a capability they had already weighted as more important, and examined the evidence behind its differing ratings. This was a change in where to investigate, not necessarily a newly discovered preference. The small historical deployment does not establish how most shoppers behave or what an AI agent would have done.
The distinction matters because an important criterion can be unhelpful for separating the options currently on the table. If every candidate meets it, another comparison may deserve the next few minutes. Deciding where to look depends on what has already been found, not just on an initial description of the desired purchase.
Keeping the work is also a cost
None of this makes continuing involvement an automatic benefit. Someone must choose the next operation and inspect the result. For the Comet user who wanted to work on something else, an assistant that repeatedly waited for another instruction could preserve exactly the burden they were trying to remove.
Elapsed duration and released attention can favor different arrangements. A slower unattended operation may be worthwhile when the person can genuinely leave it alone. Neither benefit follows just from counting clicks or calling the system an agent.
For a reader building the literature sheet, the consequence is concrete: getting the citation into place and deciding what to write beside it need not come as one package. Browser automation can finish the transfer while leaving the next thought to its reader.
Sources
- Cling: Supporting Cross-Page Coordination with Human-Directed, In-Situ, and Small-Step Automation, October 19, 2026 preprint.
- Raul Pacheco-Vega’s literature-review spreadsheet account, June 17, 2016.
- Mesh: Scaffolding Comparison Tables for Online Decision Making, UIST 2020, field deployment.
- Comet users’ discussion of Zoho setup and attention, personal accounts.