Zhao, Yue, Zhang, Angela Huyue, Yang, S. Alex, Wei, Hua, Hu, Xiyang · Nov 13, 2026
What LLM Agents See and What Users Say
Markdown · JSON · Original source
The authors of “What Agents See and What Users Say” study when an LLM agent should pause before acting on a user’s behalf. They frame the problem through presupposition theory: claims an action depends on must be supported by what the agent sees, answered by the user, or disclosed while the action can still be undone. Their experiments suggest that a tool’s view of competing records matters: showing a count of another match often did not stop earlier agents from acting on the visible record, while listing both records usually did. A user’s statement can also change behavior: saying a folder is empty stopped agents about as often when the tool counted its one item as when it listed the item, and false statements mostly reduced action. In a retail with two customers sharing a name and zip code, returning the wrong customer first led to wrong-account writes in about a quarter of conversations; checking a street the user knows kept completion near the one-customer level. The authors argue that safer writes need lookups that reveal or check competing records, and evaluations that vary what tools show. The abstract reports these study results and recommendations, but does not give the full experimental setup or enough detail to assess how broadly the findings apply.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 13, 2026 (2026-11-13).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-16T06:02:49.046Z. First observed: 2026-11-16T05:06:31.667Z. Retrieved: 2026-11-16T06:01:59.609Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The available evidence is the paper’s abstract; it does not provide the full experimental setup or enough detail to assess how broadly the reported findings apply.
Open question
Do the reported effects of listing or checking competing records hold across other tasks and tool designs?
Source-stated
The authors propose that before acting, an agent should establish the claims its action depends on through tool evidence, a user answer, or disclosure while the action remains undoable.
an agent may act only if each claim behind the act (e.g., one reservation fits the name) is seen to hold, answered by the user, or disclosed while the act can be undone
Source-stated
In the reported comparison, listing both matches stopped most earlier agents, while a count of a second match left most acting on the visible record.
a count of a second record leaves most of 36 earlier agents acting on the one they see, whereas listing both stops most
Source-stated
When users said a folder was empty, a count of its one item stopped agents about as often as listing it; stating the claim mostly reduced action when it was false.
a count of its one item stops agents about as often as a listing, and stating the claim mostly cuts acting when it is false
Source-stated
In the retail benchmark, returning a same-name, same-zip customer first led to wrong-account writes in about a quarter of conversations; checking a street the user knows kept completion near the one-customer level.
a lookup returning that customer first led to wrong-account writes in about a quarter of conversations, and one that checks a street the user knows kept completion near the one-customer level
Source-stated
The authors recommend lookups that list or check competing records and evaluations that vary what tools show.
Safe writes therefore need lookups that list or check competing records, and evaluations that vary what tools show.
Source and contributor context
Zhao, Yue, Zhang, Angela Huyue, Yang, S. Alex, Wei, Hua, Hu, Xiyang · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 968d4a4d-758f-4986-8b2e-e111c17f09ed.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T06:02:49.046Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: e41dce29a23f0a14822a33bbd09d355079062d695a6ecdb61ad3e99e7cd77031. Source text hash: e94e1d0b198f3667e63caae6cd758f17f261fa47518e99ec35376038a8b18875.
Revision and source-check history
Version 1 · 2026-11-16T06:02:49.046Z
What LLM Agents See and What Users Say
The authors of “What Agents See and What Users Say” study when an LLM agent should pause before acting on a user’s behalf. They frame the problem through presupposition theory: claims an action depends on must be supported by what the agent sees, answered by the user, or disclosed while the action can still be undone. Their experiments suggest that a tool’s view of competing records matters: showing a count of another match often did not stop earlier agents from acting on the visible record, while listing both records usually did. A user’s statement can also change behavior: saying a folder is empty stopped agents about as often when the tool counted its one item as when it listed the item, and false statements mostly reduced action. In a retail with two customers sharing a name and zip code, returning the wrong customer first led to wrong-account writes in about a quarter of conversations; checking a street the user knows kept completion near the one-customer level. The authors argue that safer writes need lookups that reveal or check competing records, and evaluations that vary what tools show. The abstract reports these study results and recommendations, but does not give the full experimental setup or enough detail to assess how broadly the findings apply.
Original source
cs.AI · Article
Source details
Original source date: Nov 13, 2026 (2026-11-13).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-16T06:02:49.046Z. First observed: 2026-11-16T05:06:31.667Z. Retrieved: 2026-11-16T06:01:59.609Z.
Limits
The available evidence is the paper’s abstract; it does not provide the full experimental setup or enough detail to assess how broadly the reported findings apply.
Open question
Do the reported effects of listing or checking competing records hold across other tasks and tool designs?
Source-stated
The authors propose that before acting, an agent should establish the claims its action depends on through tool evidence, a user answer, or disclosure while the action remains undoable.
an agent may act only if each claim behind the act (e.g., one reservation fits the name) is seen to hold, answered by the user, or disclosed while the act can be undone
Source-stated
In the reported comparison, listing both matches stopped most earlier agents, while a count of a second match left most acting on the visible record.
a count of a second record leaves most of 36 earlier agents acting on the one they see, whereas listing both stops most
Source-stated
When users said a folder was empty, a count of its one item stopped agents about as often as listing it; stating the claim mostly reduced action when it was false.
a count of its one item stops agents about as often as a listing, and stating the claim mostly cuts acting when it is false
Source-stated
In the retail benchmark, returning a same-name, same-zip customer first led to wrong-account writes in about a quarter of conversations; checking a street the user knows kept completion near the one-customer level.
a lookup returning that customer first led to wrong-account writes in about a quarter of conversations, and one that checks a street the user knows kept completion near the one-customer level
Source-stated
The authors recommend lookups that list or check competing records and evaluations that vary what tools show.
Safe writes therefore need lookups that list or check competing records, and evaluations that vary what tools show.
Source and contributor context
Zhao, Yue, Zhang, Angela Huyue, Yang, S. Alex, Wei, Hua, Hu, Xiyang · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 968d4a4d-758f-4986-8b2e-e111c17f09ed.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T06:02:49.046Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: e41dce29a23f0a14822a33bbd09d355079062d695a6ecdb61ad3e99e7cd77031. Source text hash: e94e1d0b198f3667e63caae6cd758f17f261fa47518e99ec35376038a8b18875.
uncertain · 2026-11-16T06:02:49.046Z · anthropic/claude-opus-5.5
Publication: 968d4a4d-758f-4986-8b2e-e111c17f09ed
attribution — pass: The findings and recommendations are attributed to the paper's authors. The arXiv listing identifies them as Yue Zhao, Angela Huyue Zhang, S. Alex Yang, Hua Wei and Xiyang Hu. The draft does not present the results as Escape Meridian's own conclusions.
Authors:Yue Zhao, Angela Huyue Zhang, S. Alex Yang, Hua Wei, Xiyang Hu
summary — uncertain: The summary leaves out the abstract's finding that the newest Claude and GPT models always open a counted second record and then never act. Without it, the phrase 'earlier agents' is unexplained, and readers may think the count-versus-list weakness applies to current frontier models. That changes the reading's significance. The summary also omits the stated bound: when lookups cannot tell one record from two, agents act unasked equally often in both cases. Smallest repair: add one sentence on the newest models and say that the 36 agents tested were earlier models.
The newest Claude and GPT models always open a counted second record, then never act.
date — pass: The originalPublished date of 2026-11-13 matches the v1 submission date on arXiv. It is not the later feed date.
[Submitted on 13 Nov 2026]
support — pass: Each claim is quoted from the abstract and paraphrased faithfully. The summary's phrase 'false statements mostly reduced action' slightly compresses the source, which says stating the claim mostly cuts acting when the claim is false, but the meaning is not materially changed. The question is open-ended and follows from the evidence. The limitation correctly notes that only the abstract is available.
stating the claim mostly cuts acting when it is false
media — pass: The selected media item is the verified arXiv abstract page for this paper. It is the direct original source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A language model with many learned numerical settings, used to process and generate language.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA language model with many learned numerical settings, used to process and generate language.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv