The Self-Merge Illusion in Agentic Software Engineering: Separating Developer Proxy Commits from Maintainer Acceptance across 332,000 Pull Requests
Year: 2026
Type: article
Abstract: Background: In Software Engineering 3.0, raw differences in pull request (PR) merge rates across autonomous coding agents have often been interpreted as an "autonomy acceptance premium," assuming aggregate merge rates reflect maintainer trust. Methods: We analyze 361,296 pull requests (332,406 closed PRs) across 25,402 GitHub repositories from the AIDev ecosystem (March 31, 2026 cutoff, commit c81e9b1). We specify multivariable logistic regressions clustered by repository (N = 331,945), isolate human-reviewed PRs (N = 74,997), estimate within-repository dual-use linear probability models across 2,416 repositories (N = 91,489), and fit censored Cox Proportional Hazards models (N = 331,945). Results: (1) Self-Merge Disparity: 59.22% of all merged agent PRs (146,138 / 246,777) are self-merges by the author, and 26.83% (66,215) are merged by the repository owner. Developer-proxy tools exhibit extreme author self-merging (Codex: 92.40%, median 0.23h; Cursor: 86.60%; Claude Code: 77.57%). In contrast, GitHub Copilot records 0.00% author self-merges by definitional construction due to bot-account authoring, yet owner-merges exhibit close parity (Copilot: 22.89%; Claude Code: 24.68%; Devin: 0.09% author self-merges, 6.14% owner-merges). (2) Peer-Review Equalizer: In pooled regressions, Claude Code shows an apparent advantage (OR = 2.033, p = 2.42e-25). Under human peer review, this advantage attenuates (OR = 1.023 [0.858, 1.220], p = 0.8017). (3) Within-Repo Dual-Use: In repositories using both tools, Claude Code leads unreviewed PRs by +17.72 pp via self-merges; in the human-reviewed subsample (>= 1 human review), Claude Code holds a +2.00 pp gap [0.10, 3.90], which rules out an advantage larger than approximately 5 pp and is non-significant after False Discovery Rate control (p = 0.0400, threshold q < 0.01). (4) Survival: Proxy tools resolve rapidly (Codex HR = 2.430; Claude Code HR = 1.340), while autonomous teammates exhibit similar turnaround without hazard acceleration relative to Copilot (Devin HR = 0.971, p = 0.4485; time until 50% merged under cumulative incidence: Devin 24.09h vs. Copilot 25.39h). (5) Placebo: A 1,000-iteration permutation placebo test centered at zero (mean delta: 0.0006 pp) confirms empirical robustness against the observed +15.77 pp gap (p = 0.0010). Conclusions: The autonomy acceptance premium is largely an artifact of unreviewed developer self-merges. Under human peer review, agent acceptance rates converge, ruling out substantial autonomy premiums. Replication Package: Included in the record (aidev-autonomous-coding-agents-reproducibility.tar.gz) and on GitHub at https://github.com/ZannixLabs/aidev-autonomous-coding-agents.
Author: Baba, Ajmal
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225581045