arXiv:2604.18334v1 [cs.SE] 20 Apr 2026
Reliability of AI Bots Footprints in GitHub Actions CI/CD Workflows Syed Muhammad Ashhar Shah
Sehrish Habib
Muizz Hussain
[email protected] Lahore University of Management Sciences Lahore, Pakistan
[email protected] Lahore University of Management Sciences Lahore, Pakistan
[email protected] Lahore University of Management Sciences Lahore, Pakistan
Maryam Abdul Ghafoor
Abdul Ali Bangash
[email protected] Lahore University of Management Sciences Lahore, Pakistan
[email protected] Lahore University of Management Sciences Lahore, Pakistan
Abstract Continuous Integration and Deployment (CI/CD) workflows are central to modern software delivery, yet the reliability of agentic AI bots operating within these workflows remain underexplored. Using pull requests (PRs), commits, and repositories from the AIDev dataset, we retrieved associated CI/CD workflow runs via the GitHub Actions API and analyzed 61,837 runs from 2,355 repositories, all triggered by PRs generated by five AI bots: Claude, Devin, Cursor, Copilot, and Codex. We observed substantial agentdependent differences in workflow reliability, with Copilot and Codex achieving the highest success rates ∼93% and ∼94% respectively. At the repository level, we find a negative correlation between AI agent contribution frequency and workflow success rate, suggesting that a higher frequency of Agentic PRs may hinder CI/CD workflow reliability. We defined a taxonomy of 13 categories against 3,067 agentic PRs whose associated workflows failed, and observed a trend analysis that indicates visually observable shifts from functional to non-functional PR categories over time, although these trends are not statistically significant. Our findings motivate the need for actionable guidance on integrating AI agents into CI/CD workflows and prioritizing safeguards in workflows where failures are most likely to occur. ACM Reference Format: Syed Muhammad Ashhar Shah, Sehrish Habib, Muizz Hussain, Maryam Abdul Ghafoor, and Abdul Ali Bangash. 2026. Reliability of AI Bots Footprints in GitHub Actions CI/CD Workflows. In 23rd International Conference on Mining Software Repositories (MSR ’26), April 13–14, 2026, Rio de Janeiro, Brazil. ACM, New York, NY, USA, 5 pages. https://doi.org/10.1145/3793302.3793569
1
Introduction
Continuous Integration and Continuous Deployment (CI/CD) systems automate the build, test, and deployment stages of software workflows [14, 17]. The rise of autonomous code-generating agents
This work is licensed under a Creative Commons Attribution 4.0 International License. MSR ’26, Rio de Janeiro, Brazil © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2474-9/2026/04 https://doi.org/10.1145/3793302.3793569
has ushered in Software Engineering 3.0 (SE3.0) [21], where AI systems collaborate with developers across the entire software lifecycle. Tools such as Copilot, Codex, Claude, Cursor, and Devin use large language models (LLMs) to generate and modify code [4, 7]. Several agents now operate with near end-to-end autonomy, from interpreting software issues to submitting patches and deploying fixes [16, 21]. Despite rapid adoption, we still know little about how these AI agents behave inside CI/CD environments [3]. Existing work has studied LLMs for code generation, testing, debugging, and documentation [8, 19]. However, the reliability of CI/CD workflows triggered by agentic pull requests, which we define in this study as the CI workflow success rate measured by the proportion of workflow runs that complete successfully, remains largely unexamined. We address this knowledge gap through a large-scale empirical study of 61,837 GitHub Actions workflow executions triggered by 3,067 AI-agent PRs across open-source repositories. Our study is guided by two research questions: RQ1: How does Agentic AI contributions affect CI/CD workflow outcomes? RQ2: Among Agentic PRs that lead to workflow failures, which PR types are most prevalent, and how do they evolve over time? Our work makes two primary contributions: (i) a quantitative analysis of workflow success and failure across AI agents and programming language types, (ii) a characterization and trend analysis of PRs that lead to CI/CD failures. To promote open-science, we have shared all our data, scripts, and experimental artifacts in a replication package on Zenodo [18].
2
Background and Related Work
Recent work explored how AI integrates into software engineering tasks, with several studies analyzing its impact on code generation, debugging, automated testing, and documentation [8, 19]. Mining-based analysis of GitHub repositories further examine how developers collaborate with AI-powered agents, revealing patterns in AI-assisted commits, PRs, and issue discussions [14, 17]. Foundational work on CI/CD reliability has examined why workflows fail. Prior work shows that CI/CD reliability depends heavily on CI configuration and usage. Gallaba et al. [12] found widespread misuse
MSR ’26, April 13–14, 2026, Rio de Janeiro, Brazil
Syed Muhammad Ashhar Shah, Sehrish Habib, Muizz Hussain, Maryam Abdul Ghafoor, and Abdul Ali Bangash
of Travis CI features and recurring anti-patterns that lead to fragile or misleading build outcomes, indicating that CI failures often stem from configuration rather than code defects. Beller et al. [5] provides an early empirical study of Travis CI, showing how test failures propagate through automated builds. More recently, Aïdasso et al. [2] uncover “silent failures” and misleading success signals in industrial CI systems, highlighting the fragility of build reruns and reliability reporting. Parallel work on AI-assisted software engineering identifies both productivity benefits and emerging risks. Cotroneo et al. [9] conduct a large-scale comparison of humanwritten and AI-generated code, showing that AI-produced artifacts exhibit different defects, vulnerabilities, and complexity profiles. Hassan et al. [13] outline the foundational principles of agentic software engineering and articulate open research challenges related to reliability, reproducibility, and safe autonomy. In this study, we analyze 208,843 CI/CD workflow runs triggered by 33,596 PRs generated by five agentic AI bots, spanning 88,576 commits across 2,355 GitHub OSS repositories. From these, we focus on 61,837 workflows where the AI agent triggered the CI/CD workflow through a PR and identify 3,067 PRs that result in workflow failures. We use topic modeling to develop a taxonomy of 13 PR types linked to failed workflows and examine how the percentage composition of each PR type evolve over time.
3 Methodology 3.1 Data Collection We use the AIDev dataset [15], a large-scale collection of GitHub repositories curated for studying AI-driven software development in OSS GitHub repositories. The dataset provides metadata such as commits, pull requests (PRs), and issues, enabling fine-grained analysis of agentic AI bots. We load the dataset into PostgreSQL for SQL-based retrieval and transformation. We focus on the 2,807 repositories listed in the pr_repositories table of the AIDev dataset, which explicitly excludes toy, inactive, or artificially generated projects. To retrieve the relevant GitHub Actions workflow files, we search for YAML files (.yml) within each repository’s .github/workflows directory. After this filtering step, 2,355 repositories remain for further analysis. Furthermore, from the pull_request and pr_commits tables of the AIDev dataset, we extract 33,596 pull requests and 88,576 commits corresponding to repositories that contain at least one valid CI/CD workflow. Each PR has multiple commits, we distribute these commits into commit-PR pairs. For each commit–PR pair, we collect workflow execution metadata, including trigger type, which denotes the GitHub event that initiated the workflow, job status, which indicates the final execution outcome (e.g., success or failure), and duration, which measures the total time taken to complete the workflow, via the GitHub Actions API, resulting in 208,843 workflow runs. To complement the AIDev dataset, we filter these runs to those initiated by the PRs made by the five agentic AI bots (Claude, Devin, Cursor, Copilot, and Codex), yielding 71,241 workflow runs. We further restrict our analysis to workflows that have either succeeded or failed, producing a final set of 61,837 workflow runs that we use in our quantitative and qualitative analysis.
Figure 1 presents the end-to-end pipeline of our methodology, i.e., from data extraction and repository filtering to workflow metadata retrieval and analysis.
3.2
Language type tagging
To understand the success of Agentic PRs in the CI/CD workflows, we categorize the PRs into their repository’s language type (highlevel language or low-level language). To figure out the language of a repository, we use the pr_repository table of the AIDev dataset which has this information. Following the standard convention [1], we group languages into low-level (C, C++, C#, Rust, Zig, Assembly) and high-level (Python, Java, JavaScript, TypeScript, Ruby, PHP, Go, Kotlin, Swift, R, Dart, HTML, CSS, Shell, and others). After this categorization, we tag 61,837 workflows with a language type (i.e., the language used in the repository of their PR), we find that 53,087 workflows use high-level languages, 8,192 workflows use low-level languages, and the remaining 558 had no language specified.
3.3
Workflow success rate
For each repository, we compute a workflow success rate by dividing the number of successful workflow runs by the number of total workflow runs. We then examine how the workflow success rate varies with the frequency of agentic PRs, enabling us to assess whether higher levels of AI contribution are associated with CI/CD workflow success/failure rates.
3.4
PRs Categorization
We categorize the PRs that trigger failing CI/CD workflows. In our dataset, 3,067 PRs trigger CI/CD workflows that fail to resolve. To categorize PRs, we sample 548 PRs from 3,067 PRs (with 99% confidence level and a 5% margin of error) and perform initial semantic clustering on the PR titles using GPT 5.0 [22]. Two authors then refine these clusters through manual sorting, resolving disagreements via a negotiated agreement process [6]. This process produced final categories of 13 PR types: Bug Fixes, Testing & Quality Assurance, New Features & Enhancements, APIs, SDKs & Integrations, User Interface & User Experience, Configuration & Infrastructure, Refactoring & Code Quality, Documentation & Examples, Security & Authentication, Performance & Optimization, Regular Maintenance & Miscellaneous, Tools, Utilities & CLI, and Continuous Integration & Continuous Deployment. To perform close-card sorting, we then provide GPT 5.0 the remaining PRs’ titles and the finalized 13 PR categories from the previous step. The model then assigned each PR to the most relevant category. To further validate the labeling of PRs, we manually sample 334 PRs out of 2,519 PRs (with 95% confidence level, 5% margin of error). Two authors independently coded the samples, and we found a strong inter-rater agreement between human and GPT 5.0, through a Cohen’s Kappa of 0.88. This step resulted in a labeled dataset of 3,067 PRs with assigned PR categories.
Reliability of AI Bots Footprints in GitHub Actions CI/CD Workflows
MSR ’26, April 13–14, 2026, Rio de Janeiro, Brazil
Have GitHub Action Workflows?
2,807 Repositories
Fetch Associated PRs
2,355 Repositories
1
2
Fetch Associated Commits
For Each Commit-PR Pair, Fetch Workflow Metadata From GitHub Actions API
Generate Categories & Label PR’s
Triggering Actor = Claude, Devin, Cursor, Copilot, OpenAI
Filter pullrequests where workflows failed
7
6
208,843 Workflow Runs 3
88,576 Commits Against Pull Requests
33,596 Pull Requests
GPT 5.0 & Human Categorization
3,067 Pull Requests
PR’s With Assigned Categories
5
4
71,241 Workflow Runs
Workflows that either failed or were sucessful
61,837 Workflow Runs
2519 Pull Request Titles With Categories RQ2 Analysis
RQ1 Analysis
Figure 1: Overview of the CI/CD Reliability Analysis Workflow Table 1: Agent Success Rate Distribution by Language Level Agent Claude Copilot Cursor Devin Codex
High-Level Success Rate Total Runs 63.89% 36.00 93.10% 9,349.00 73.29% 2,696.00 78.03% 40,921.00 92.94% 85.00
Low-Level Success Rate Total Runs 100.00% 1.00 93.62% 4,830.00 65.07% 335.00 69.05% 2,931.00 95.79% 95.00
Total Success Rate Total Runs 64.86% 37.00 93.28% 14,179.00 72.39% 3,031.00 77.43% 43,852.00 94.44% 180.00
4 Results 4.1 CI/CD Reliability Under Increasing AI Contributions (RQ1) 4.1.1 Cross Agent and Language Type Reliability. In terms of workflow success rate, there is a significant difference between the agents, both for high-level language contributions and low-level language contributions. We find that Copilot and Codex achieves the highest success at 93.28% and 94.44%, while Claude had the lowest reliability at 64.86%. Devin exhibits moderate reliability at 77.43%, despite contributing most (71.56%) of the workflows runs. We detail these stats in Table 1, A chi-square test indicates statistically significant differences in workflow success rates across all agents (𝑝 < 0.01, 𝐶𝑟𝑎𝑚𝑒𝑟 ′𝑠 𝑉 = 0.177). Pairwise post-hoc Fisher’s exact tests with Benjamini–Hochberg correction further confirm these differences. Copilot has significantly higher odds of workflow success than Claude (OR = 7.53), Cursor (OR = 5.29), and Devin (OR = 4.05), while showing no statistically significant difference relative to Codex after correction (OR = 0.816). Here, an Odds Ratio (OR) greater than 1 indicates that the first agent has higher odds of success than the second, while OR less than 1 indicates lower odds. Our analysis in the context of programming language types show that agents have a slightly higher CI/CD reliability success on PRs for low-level languages (86.7%) compared to high-level languages (82.4%). A chi-square and effect size test confirms that this difference is statistically significant with a small effect size (𝑝 < 0.01, 𝐶𝑟𝑎𝑚𝑒𝑟 ′𝑠 𝑉 = 0.028). The small effect size
Figure 2: Scatter plot of Agentic PRs vs. workflow success rate per repository
shows programming language might not be the major factor for cross agent differences. 4.1.2 Relation of Agent PR Frequency with Workflow Outcomes. We examine the relationship between AI agents contribution frequency and CI/CD workflow success rate at the repository level. Figure 2 presents a logarithmic scatter plot of the number of Agentic PRs per repository against the corresponding workflow success rate. The results show a weak negative corelation between Agent PR frequency and workflow success rate (Spearman 𝜌 = −0.34, 𝑝 < 0.01), indicating that repositories with more Agentic PRs tend to exhibit lower workflow success rates. This suggests that increased agentic activity may degrade CI/CD performance.
4.2
Temporal Trends of PR Categories (RQ2)
In RQ2, we analyze the PR category-wise differences and inspect, through a trend analysis, how they evolve over-time. Specifically, we look at 2,519 PRs that triggered 9,012 failed workflows.
MSR ’26, April 13–14, 2026, Rio de Janeiro, Brazil
Syed Muhammad Ashhar Shah, Sehrish Habib, Muizz Hussain, Maryam Abdul Ghafoor, and Abdul Ali Bangash
assign each PR to a single category for analysis, even though some PRs could span multiple types of changes. Conclusion Validity. Our statistical analysis assume that workflow runs are independent. In practice, repositories might contain multiple related workflows, introducing potential dependencies. Future work could employ hierarchical or mixed-effects models to account for these nested executions more accurately.
6
Figure 3: Month-by-Month Trends in PR Categories Leading to Workflow Failures (Dec 2024 –Jul 2025)
Failure Category Distribution. Workflow failures are unevenly distributed across PR types, with higher proportions observed for Bug Fixes (17.57%), UI/UX (11.64%), New Features & Enhancements (10.26%), Refactoring (10.04%), and Configuration & Infrastructure (8.04%). In contrast, categories such as Security & Authentication (3.59%), CI/CD (3.42%), and Tools & CLI (1.95%) account for substantially smaller shares. A chi-square goodness-of-fit test confirms that workflow failures are unevenly distributed across PR categories (𝑝 < 0.01), indicating failures are not uniformly spread across different contribution types. Temporal Trends in Failure Proportions. The heatmap in Figure 3 shows evolving patterns in the PR contribution types that trigger workflow failures, with each cell representing the % of PRs for that month. Functional categories, such as Bug Fixes and New Features & Enhancements, consistently account for a large share of failures, with recurring peaks (e.g., Bug Fixes in Dec 2024 and Mar 2025). Non-functional categories, such as Documentation & Examples and UI/UX, show visually increasing patterns from May to July 2025. However, Mann–Kendall trend tests on the PR categories (with Benjamini–Hochberg correction) reveals no statistically significant monotonic trends across 8 months. While Performance & Optimization shows the strongest positive signal (highest Kendall’s 𝜏), its still not significant, indicating emerging but unconfirmed shifts in CI/CD failures.
5
Threats to Validity
Internal Validity. Our analysis uses PR titles to classify the types of PRs associated with workflow failures. PR categories were generated using GPT-5.0, which may sometimes misclassify PRs. To mitigate this risk, we conducted manual human validation and observed strong inter-rater agreement between human annotations and the AI model (Cohen’s 𝜅 = 0.88, indicating high consistency in PR categorization). External Validity. Our dataset consists mainly of public GitHub repositories. Consequently, the findings might not fully generalize to private or enterprise CI/CD environments, which typically involve more complex workflows and processes. Construct Validity. We attribute agent activity based on commit authorship metadata; however, hybrid human–AI edits within the same PR might introduce minor inaccuracies. Additionally, we
Implications
For software developers and maintainers, the success rate of CI/CD workflows triggered by Agentic PRs varies across agents. Copilot and Codex generally perform well, whereas Cursor and Claude cause failures more frequently. Developers can use these insights to implement AI-aware validation gates and automated review processes to improve CI/CD workflow reliability. For CI/CD engineers, detecting and managing agentic AI commits separately can enable early anomaly detection. Our findings can help engineers prioritize monitoring PR types that more frequently trigger workflow failures and design CI/CD configurations to handle AI-generated code failures more effectively. For AI model developers and researchers, our findings identify the types of PRs that most frequently lead to workflow failures, pinpointing areas where AI agents face challenges in CI/CD workflows. By leveraging these insights, to fine-tune models on CI/CD-specific data, developers can enhance model reliability and foster safer, more robust AI coding assistants [10, 11, 20].
7
Ethical Considerations
We used the AIDev dataset and cited its authors throughout the paper to give appropriate credit.
8
Conclusion
This study provides a large-scale empirical analysis of 61,837 GitHub Actions workflow runs to evaluate the reliability of agentic AI bots’ contributions in CI/CD workflows. Our findings reveal that reliability is primarily agent-dependent: while Copilot and Codex achieve success rates exceeding 93%, other agents like Claude and Cursor show a significantly higher tendency for failure. We also observe a weak negative correlation between agentic contribution frequency and workflow success rate, suggesting that increased agentic activity may lead to more failures in CI/CD workflows. Our analysis of 3,067 PRs by their categories shows that agentic PRs are causing frequent CI/CD failures in functional categories, such as Bug Fixes. Over time, however, failures in non-functional categories, such as UI/UX and Performance & Optimization, also exhibit a growing trend, even though these increases are not statistically significant. Our results highlight the need for AI-aware validation gates and targeted monitoring of both functional and non-functional AI-based code changes. Future work will investigate the causes of agent-specific performance gaps and extend this analysis to private enterprise code bases to advance the understanding of Software Engineering 3.0.
References [1] Alfred V. Aho, Monica S. Lam, Ravi Sethi, and Jeffrey D. Ullman. 2006. Compilers: Principles, Techniques, and Tools. Pearson, Boston, MA.
Reliability of AI Bots Footprints in GitHub Actions CI/CD Workflows
[2] Henri Aïdasso, Francis Bordeleau, and Ali Tizghadam. 2025. On the Illusion of Success: An Empirical Study of Build Reruns and Silent Failures in Industrial CI. arXiv:2509.14347 [cs.SE] https://arxiv.org/abs/2509.14347 [3] Mohammad Baqar, Saba Naqvi, and Rajat Khanda. 2025. AI-Augmented CI/CD Pipelines: From Code Commit to Production with Autonomous Decisions. arXiv:2508.11867 [cs.SE] https://arxiv.org/abs/2508.11867 [4] Shraddha Barke, Michael B. James, and Nadia Polikarpova. 2022. Grounded Copilot: How Programmers Interact with Code-Generating Models. arXiv:2206.15000 [cs.HC] https://arxiv.org/abs/2206.15000 [5] Moritz Beller, Georgios Gousios, and Andy Zaidman. 2017. Oops, My Tests Broke the Build: An Explorative Analysis of Travis CI with GitHub. In 2017 IEEE/ACM 14th International Conference on Mining Software Repositories (MSR). 356–367. doi:10.1109/MSR.2017.62 [6] Virginia Braun and Victoria Clarke. 2006. Using thematic analysis in psychology. Qualitative Research in Psychology 3 (01 2006), 77–101. doi:10.1191/ 1478088706qp063oa [7] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. 2021. Evaluating Large Language Models Trained on Code. arXiv:2107.03374 [cs.LG] https://arxiv.org/abs/2107.03374 [8] Moataz Chouchen, Narjes Bessghaier, Mahi Begoug, Ali Ouni, Eman Alomar, and Mohamed Wiem Mkaouer. 2024. How Do Software Developers Use ChatGPT? An Exploratory Study on GitHub Pull Requests. In Proceedings of the 21st International Conference on Mining Software Repositories (Lisbon, Portugal) (MSR ’24). Association for Computing Machinery, New York, NY, USA, 212–216. doi:10.1145/3643991.3645084 [9] Domenico Cotroneo, Cristina Improta, and Pietro Liguori. 2025. Human-Written vs. AI-Generated Code: A Large-Scale Study of Defects, Vulnerabilities, and Complexity. arXiv:2508.21634 [cs.SE] https://arxiv.org/abs/2508.21634 [10] Matheus De Morais Leça, Mariana Bento, and Ronnie De Souza Santos. 2025. Responsible AI in the Software Industry: A Practitioner-Centered Perspective. In 2025 IEEE/ACM International Workshop on Responsible AI Engineering (RAIE). 37–40. doi:10.1109/RAIE66699.2025.00011 [11] Ahmad Fadavi and Mohsen Alizadeh. 2024. Ownership of Artificial IntelligenceGenerated Works: An Overview of the Emerging Intellectual Property Challenges in the Technology Era. (11 2024). doi:10.30497/law.2024.245828.3492 [12] Keheliya Gallaba and Shane McIntosh. 2018. Use and Misuse of Continuous Integration Features: An Empirical Study of Projects That (Mis)Use Travis CI. IEEE Transactions on Software Engineering PP (05 2018), 1–1. doi:10.1109/TSE. 2018.2838131 [13] Ahmed E. Hassan, Hao Li, Dayi Lin, Bram Adams, Tse-Hsun Chen, Yutaro Kashiwa, and Dong Qiu. 2025. Agentic Software Engineering: Foundational Pillars and a Research Roadmap. arXiv:2509.06216 [cs.SE] https://arxiv.org/abs/ 2509.06216 [14] Michael Hilton, Timothy Tunnell, Kai Huang, Darko Marinov, and Danny Dig. 2016. Usage, costs, and benefits of continuous integration in open-source projects. In Proceedings of the 31st IEEE/ACM International Conference on Automated Software Engineering (Singapore, Singapore) (ASE ’16). Association for Computing Machinery, New York, NY, USA, 426–437. doi:10.1145/2970276.2970358 [15] SAIL Research. 2026. AIDev Challenge Dataset: AI Teammates in Software Engineering 3.0. https://github.com/SAILResearch/AI_Teammates_in_SE3/blob/ main/AIDev_Challenge.pdf. Accessed: 2025-10-04. [16] Sakshi Kini Sana Ansari. 2024. The World’s First AI Software Engineer, DEVIN AI. INTERNATIONAL JOURNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY (2024). [17] Jadson Santos, Daniel Alencar da Costa, and Uirá Kulesza. 2022. Investigating the Impact of Continuous Integration Practices on the Productivity and Quality of Open-Source Projects. In Proceedings of the 16th ACM / IEEE International Symposium on Empirical Software Engineering and Measurement (ESEM ’22). ACM, 137–147. doi:10.1145/3544902.3546244 [18] Syed Muhammad Ashhar Shah, Sehrish Habib, Muizz Hussain, Maryam Abdul Ghafoor, and Abdul Ali Bangash. 2025. Reliability of AI Bots Footprints in GitHub Actions CI/CD Workflows - Replication Package. doi:10.5281/zenodo. 17867301 [19] Tao Xiao, Youmei Fan, Fabio Calefato, Christoph Treude, Raula Gaikovina Kula, Hideaki Hata, and Sebastian Baltes. 2025. Self-Admitted GenAI Usage in OpenSource Software. arXiv:2507.10422 [cs.SE] https://arxiv.org/abs/2507.10422
MSR ’26, April 13–14, 2026, Rio de Janeiro, Brazil
[20] Alan Yang and T. Andrew Yang. 2024. Social Dangers of Generative Artificial Intelligence: Review and Guidelines. In Proceedings of the 25th Annual International Conference on Digital Government Research (Taipei, Taiwan) (dg.o ’24). Association for Computing Machinery, New York, NY, USA, 654–658. doi:10.1145/3657054.3664243 [21] John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. 2024. SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering. arXiv:2405.15793 [cs.SE] https://arxiv.org/abs/2405.15793 [22] Zhimin Zhao, Yihao Chen, Abdul Ali Bangash, Bram Adams, and Ahmed E. Hassan. 2024. An empirical study of challenges in machine learning asset management. Empirical Software Engineering 29, 4 (June 2024). doi:10.1007/s10664024-10474-4