Rahman, Musfiqur, Begoug, Mahi, Shihab, Emad · Nov 16, 2026
Rahman, Begoug and Shihab: Revisiting the Survival of Agent-Attributed Code
Markdown · JSON · Original source
Authors Musfiqur Rahman, Mahi Begoug and Emad Shihab revisit whether code from AI coding agents remains in software after it is merged. They analyze 854,139 lines added by 2,802 merged pull requests across 215 AIDev repositories, varying nine measurement choices in 88 survival analyses. Their results differ by agent type: for human-supervised agents, mostly GitHub Copilot, choices change the effect’s size but not its direction; for Devin, they can change the verdict. Treating lines as independent, 15 of 22 specifications report Devin’s code modified sooner; clustering standard errors by repository reduces that count to 3, no more than permuted labels produce. Within repositories, no specification detects a difference. The authors also report that humans make most modifications to agent code under every unit and observation window. Their central argument is that conclusions about whether agent code survives longer than human code depend substantially on measurement choices, and they derive an operationalization checklist for post-merge studies.
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 16, 2026 (2026-11-16).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-18T13:17:44.623Z. First observed: 2026-11-18T05:22:21.067Z. Retrieved: 2026-11-18T13:16:52.931Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The evidence provided is the paper’s abstract; it does not include the full methods, data, or operationalization checklist. The findings described concern the AIDev dataset and the specifications reported in the abstract; they do not establish how results would compare in other datasets.
Open question
Would the reported difference for Devin persist under other measurement choices or datasets beyond the AIDev repositories studied?
Source-stated
The authors revisit code survival using 854,139 lines from merged pull requests in 215 repositories, and vary measurement choices across 88 survival analyses.
854,139 lines that 2,802 merged pull requests (PRs) added to 215 repositories of the AIDev dataset. We vary nine measurement choices
Source-stated
For code from human-supervised agents, mostly GitHub Copilot, measurement choices alter effect size but not direction; for Devin, choices decide the verdict.
For code from human-supervised agents, mostly GitHub Copilot, the choices change the size of the effect but not its direction. For code from the fully-autonomous agent Devin, they decide the verdict.
Source-stated
The count of specifications reporting Devin code modified sooner falls when standard errors are clustered by repository, and within-repository analyses detect no difference.
15 of 22 specifications report it modified sooner when lines are treated as independent, but only 3 when standard errors are clustered by repository, no more than permuted labels produce. Within repositories, no specification detects a difference.
Source-stated
The authors argue that measurement choices substantially determine conclusions about whether agent-attributed code survives longer than human code.
Whether agent code survives longer than human code is therefore decided as much by the measurement as by the code.
Source-stated
The authors report that humans make most modifications to agent code across the units and observation windows they examined.
Humans make most modifications to agent code under every unit and window.
Source and contributor context
Rahman, Musfiqur, Begoug, Mahi, Shihab, Emad · 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.
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: ee6591d0-6add-46b7-85d7-164bbc80a166.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-18T13:17:44.623Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: useful_perspective.
Observation: f32b30678a63c256f422695cda791c6871a74813b813aeba2c497ba99719809b. Source text hash: e84391b95190ab44ca1a07ae3c5a2fc9ffa4cd895525f0bfc98d49ce3bfb0de9.
Revision and source-check history
Version 1 · 2026-11-18T13:17:44.623Z
Rahman, Begoug and Shihab: Revisiting the Survival of Agent-Attributed Code
Authors Musfiqur Rahman, Mahi Begoug and Emad Shihab revisit whether code from AI coding agents remains in software after it is merged. They analyze 854,139 lines added by 2,802 merged pull requests across 215 AIDev repositories, varying nine measurement choices in 88 survival analyses. Their results differ by agent type: for human-supervised agents, mostly GitHub Copilot, choices change the effect’s size but not its direction; for Devin, they can change the verdict. Treating lines as independent, 15 of 22 specifications report Devin’s code modified sooner; clustering standard errors by repository reduces that count to 3, no more than permuted labels produce. Within repositories, no specification detects a difference. The authors also report that humans make most modifications to agent code under every unit and observation window. Their central argument is that conclusions about whether agent code survives longer than human code depend substantially on measurement choices, and they derive an operationalization checklist for post-merge studies.
Original source
cs.AI · Article
Source details
Original source date: Nov 16, 2026 (2026-11-16).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-18T13:17:44.623Z. First observed: 2026-11-18T05:22:21.067Z. Retrieved: 2026-11-18T13:16:52.931Z.
Limits
The evidence provided is the paper’s abstract; it does not include the full methods, data, or operationalization checklist. The findings described concern the AIDev dataset and the specifications reported in the abstract; they do not establish how results would compare in other datasets.
Open question
Would the reported difference for Devin persist under other measurement choices or datasets beyond the AIDev repositories studied?
Source-stated
The authors revisit code survival using 854,139 lines from merged pull requests in 215 repositories, and vary measurement choices across 88 survival analyses.
854,139 lines that 2,802 merged pull requests (PRs) added to 215 repositories of the AIDev dataset. We vary nine measurement choices
Source-stated
For code from human-supervised agents, mostly GitHub Copilot, measurement choices alter effect size but not direction; for Devin, choices decide the verdict.
For code from human-supervised agents, mostly GitHub Copilot, the choices change the size of the effect but not its direction. For code from the fully-autonomous agent Devin, they decide the verdict.
Source-stated
The count of specifications reporting Devin code modified sooner falls when standard errors are clustered by repository, and within-repository analyses detect no difference.
15 of 22 specifications report it modified sooner when lines are treated as independent, but only 3 when standard errors are clustered by repository, no more than permuted labels produce. Within repositories, no specification detects a difference.
Source-stated
The authors argue that measurement choices substantially determine conclusions about whether agent-attributed code survives longer than human code.
Whether agent code survives longer than human code is therefore decided as much by the measurement as by the code.
Source-stated
The authors report that humans make most modifications to agent code across the units and observation windows they examined.
Humans make most modifications to agent code under every unit and window.
Source and contributor context
Rahman, Musfiqur, Begoug, Mahi, Shihab, Emad · 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.
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: ee6591d0-6add-46b7-85d7-164bbc80a166.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-18T13:17:44.623Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: useful_perspective.
Observation: f32b30678a63c256f422695cda791c6871a74813b813aeba2c497ba99719809b. Source text hash: e84391b95190ab44ca1a07ae3c5a2fc9ffa4cd895525f0bfc98d49ce3bfb0de9.
uncertain · 2026-11-18T13:17:44.623Z · anthropic/claude-opus-5.5
Publication: ee6591d0-6add-46b7-85d7-164bbc80a166
attribution — pass: The reading correctly attributes the work to Musfiqur Rahman, Mahi Begoug and Emad Shihab and keeps the findings framed as the authors' reported results and argument.
Authors:Musfiqur Rahman, Mahi Begoug, Emad Shihab
summary — uncertain: The facts are faithful, but the central Devin result is hard for a non-specialist to follow. 'Clustering standard errors by repository', 'permuted labels' and 'specifications' are not explained, so readers cannot see why 15 falls to 3. Devin is not identified as the source's 'fully-autonomous agent', which is the contrast with human-supervised Copilot. The reading also omits the forceful finding that five published designs each flip with one choice. The question's 'reported difference for Devin' implies a robust difference, though the source says the choices decide that verdict. Smallest repair: briefly gloss these terms, add those two points and reframe the question.
For code from the fully-autonomous agent Devin, they decide the verdict.
date — pass: originalPublished 2026-11-16 matches the arXiv v1 submission date. The feed date of 2026-11-18 is correctly not used.
[Submitted on 16 Nov 2026]
support — pass: Each claim is directly supported by its excerpt from the abstract. The limitations correctly note that only the abstract is available and that the study's scope is the AIDev dataset. The reading adds no unsupported benefits or costs.
Whether agent code survives longer than human code is therefore decided as much by the measurement as by the code.
media — pass: The selected media is the verified arXiv abstract page for this exact paper, which is the most direct access to 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.
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv