Xu, Song · Nov 17, 2026
Song Xu on why idempotency is not recovery for agent tool actions
Markdown · JSON · Original source
Song Xu’s study examines what happens when a language-model agent’s tool call is interrupted after it may have changed an external system. Retrying can duplicate the action; assuming it succeeded can hide an action that never happened. The paper argues that idempotency is not the same as recovery: systems also need to determine what happened and resolve uncertain outcomes. In a controlled agent runtime, the full configuration combined authorization, an action ledger, target-side idempotency, reconciliation, and evidence sealing. Across tests on Windows and Linux, it recorded no repeated effects in 200 known crash-and-replay pairs, compared with 117 for each simpler baseline. Disabling reconciliation left all 100 known post-effect crash outcomes ambiguous, even though no duplicates were observed. These results describe the tested , not a general measure of deployed agents. The full runtime had about one-seventh the fail-open throughput, and injected evidence-write faults left 75 of 100 trials per platform unresolved. The study thus reports a tradeoff: suppressing duplicate actions does not by itself establish whether an interrupted action took effect or how to recover from it.
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 17, 2026 (2026-11-17).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-19T07:02:05.290Z. First observed: 2026-11-19T05:04:20.235Z. Retrieved: 2026-11-19T06:50:19.117Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The reported results come from a controlled agent runtime and , not a demonstration of reliability in deployed agents. The tests used Windows and Linux, 7,513 trials, 96 captured model responses, and 24 constructed cases; the evidence does not establish how results transfer to other settings. Injected evidence-write faults left 75 of 100 trials per platform unresolved. The abstract gives no further detail on the construction of cases or the runtime implementation.
Open question
How would the reported reliability and performance tradeoffs change in settings beyond the study’s controlled and constructed cases?
Source-stated
Song Xu’s paper studies interrupted language-model agent tool calls that may have changed an external system without confirmation; retrying risks duplication, while assuming success can conceal an action that never occurred.
When a language-model agent's tool call is interrupted, the action may already have changed an external system without confirmation. Retrying risks duplication; assuming success may conceal an action that never occurred.
Source-stated
The study compares a full runtime combining authorization, an action ledger, target-side idempotency, reconciliation, and evidence sealing with simpler configurations and ablations.
a controlled agent runtime combining authorization, an action ledger, target-side idempotency, reconciliation, and evidence sealing.
Source-stated
In 200 known original crash and replay pairs, the full configuration recorded no repeated effects, compared with 117 for each simpler baseline.
In 200 known original crash and replay pairs, the full configuration recorded 0/200 repeated effects versus 117/200 for each simpler baseline.
Source-stated
Disabling reconciliation left every tested known post-effect crash outcome ambiguous, despite no observed duplicates.
Disabling reconciliation left 100/100 known post-effect crash outcomes ambiguous despite no observed duplicates.
Source-stated
The full runtime had substantially lower throughput than the fail-open configuration, while evidence-write faults left many trials unresolved.
The full runtime had about one-seventh of fail-open throughput, while injected evidence-write faults left 75/100 trials per platform unresolved.
Source-stated
The findings distinguish duplicate suppression from recovery and report reliability and performance tradeoffs within the tested benchmark.
These findings distinguish duplicate suppression from recovery and identify their measured reliability and performance tradeoffs within the tested benchmark.
Source and contributor context
Xu, Song · 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.
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: 79b2a5c1-47ac-495e-83c1-2d6abcf93204.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-19T07:02:05.290Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b9b46c33732597928942fedcd1834ed89b1c2f958586c58052b2cf8134b80467. Source text hash: 21e90fdd04ec1f7823bbde94d9dbcff940c329f558d1569bf10172ea78d5817d.
Revision and source-check history
Version 1 · 2026-11-19T07:02:05.290Z
Song Xu on why idempotency is not recovery for agent tool actions
Song Xu’s study examines what happens when a language-model agent’s tool call is interrupted after it may have changed an external system. Retrying can duplicate the action; assuming it succeeded can hide an action that never happened. The paper argues that idempotency is not the same as recovery: systems also need to determine what happened and resolve uncertain outcomes. In a controlled agent runtime, the full configuration combined authorization, an action ledger, target-side idempotency, reconciliation, and evidence sealing. Across tests on Windows and Linux, it recorded no repeated effects in 200 known crash-and-replay pairs, compared with 117 for each simpler baseline. Disabling reconciliation left all 100 known post-effect crash outcomes ambiguous, even though no duplicates were observed. These results describe the tested , not a general measure of deployed agents. The full runtime had about one-seventh the fail-open throughput, and injected evidence-write faults left 75 of 100 trials per platform unresolved. The study thus reports a tradeoff: suppressing duplicate actions does not by itself establish whether an interrupted action took effect or how to recover from it.
Original source
cs.AI · Article
Source details
Original source date: Nov 17, 2026 (2026-11-17).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-19T07:02:05.290Z. First observed: 2026-11-19T05:04:20.235Z. Retrieved: 2026-11-19T06:50:19.117Z.
Limits
The reported results come from a controlled agent runtime and , not a demonstration of reliability in deployed agents. The tests used Windows and Linux, 7,513 trials, 96 captured model responses, and 24 constructed cases; the evidence does not establish how results transfer to other settings. Injected evidence-write faults left 75 of 100 trials per platform unresolved. The abstract gives no further detail on the construction of cases or the runtime implementation.
Open question
How would the reported reliability and performance tradeoffs change in settings beyond the study’s controlled and constructed cases?
Source-stated
Song Xu’s paper studies interrupted language-model agent tool calls that may have changed an external system without confirmation; retrying risks duplication, while assuming success can conceal an action that never occurred.
When a language-model agent's tool call is interrupted, the action may already have changed an external system without confirmation. Retrying risks duplication; assuming success may conceal an action that never occurred.
Source-stated
The study compares a full runtime combining authorization, an action ledger, target-side idempotency, reconciliation, and evidence sealing with simpler configurations and ablations.
a controlled agent runtime combining authorization, an action ledger, target-side idempotency, reconciliation, and evidence sealing.
Source-stated
In 200 known original crash and replay pairs, the full configuration recorded no repeated effects, compared with 117 for each simpler baseline.
In 200 known original crash and replay pairs, the full configuration recorded 0/200 repeated effects versus 117/200 for each simpler baseline.
Source-stated
Disabling reconciliation left every tested known post-effect crash outcome ambiguous, despite no observed duplicates.
Disabling reconciliation left 100/100 known post-effect crash outcomes ambiguous despite no observed duplicates.
Source-stated
The full runtime had substantially lower throughput than the fail-open configuration, while evidence-write faults left many trials unresolved.
The full runtime had about one-seventh of fail-open throughput, while injected evidence-write faults left 75/100 trials per platform unresolved.
Source-stated
The findings distinguish duplicate suppression from recovery and report reliability and performance tradeoffs within the tested benchmark.
These findings distinguish duplicate suppression from recovery and identify their measured reliability and performance tradeoffs within the tested benchmark.
Source and contributor context
Xu, Song · 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.
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: 79b2a5c1-47ac-495e-83c1-2d6abcf93204.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-19T07:02:05.290Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b9b46c33732597928942fedcd1834ed89b1c2f958586c58052b2cf8134b80467. Source text hash: 21e90fdd04ec1f7823bbde94d9dbcff940c329f558d1569bf10172ea78d5817d.
passed · 2026-11-19T07:02:05.290Z · anthropic/claude-opus-5.5
Publication: 79b2a5c1-47ac-495e-83c1-2d6abcf93204
attribution — pass: The page lists Song Xu as the sole author. The draft attributes the study, its framing and its measurements to Song Xu and the paper, and does not present them as established facts or Escape Meridian's own view.
Authors:Song Xu
summary — pass: The draft opens with the interrupted tool-call problem in plain terms: retrying can duplicate an action, while assuming success can hide one that never happened. That framing conveys idempotency as duplicate suppression and keeps the author's thesis that this is not recovery. The key numbers and benchmark scope are preserved. Minor: 'Across tests on Windows and Linux' leaves it unclear whether the 200 pairs are per platform, though the source is itself ambiguous. 'Fail-open throughput' would benefit from a tooltip. The question follows from the stated scope.
Retrying risks duplication; assuming success may conceal an action that never occurred.
date — pass: The originalPublished date of 2026-11-17 matches the arXiv v1 submission date. The feed date of 2026-11-19 is later and correctly not used.
[Submitted on 17 Nov 2026]
support — pass: Each claim matches the abstract's figures: 0/200 vs 117/200 repeated effects, 100/100 ambiguous outcomes without reconciliation, about one-seventh of fail-open throughput, and 75/100 unresolved trials. The limitations correctly confine results to the tested benchmark and constructed cases. The draft adds no unsupported benefits or costs.
These findings distinguish duplicate suppression from recovery and identify their measured reliability and performance tradeoffs within the tested benchmark.
media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct way to reach the 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 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 defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
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 defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference