When Agents Break Contracts: A Socio-Technical Framework for AI-Authored Pull Requests and API Breaking Changes

AI coding agents now open pull requests (PRs) on real repositories at scale, and recent mining studies show that many of these PRs are never merged for reasons that go beyond incorrect code: duplicated work, unwanted features, ignored feedback, and reviewers who never engage. In parallel, web API evolution remains a persistent source of client breakage, and mobile clients are especially exposed because users update apps at different times, so old app versions keep calling an API after it has changed. These two literatures have developed independently. This paper synthesizes both and argues that they meet at a concrete, under-studied point: agent-authored changes to API-bearing code are consequential contract-affecting changes that are reviewed by humans who may not have the client-side context needed to judge them. We (i) summarize what is known about failures of agentic PRs and about breaking-change detection and classification, (ii) propose a three-layer socio-technical framework (technical, communication, organizational) extended with a fourth, mobile-specific client-lag layer, (iii) identify five gaps in the current evidence, and (iv) lay out a research agenda with research questions, hypotheses, data sources, and an evaluation protocol for predicting breaking changes before merge. This is primarily a survey and position paper. Its only empirical component is a preliminary analysis of the AIDev corpus: 4,402 of 361,296 agent PRs (1.2%) touch API-definition files, agents differ markedly in how often they do so, and in a 150-PR sample of OpenAPI changes, 14 of the 108 that could be automatically diffed (13.0%) contain at least one breaking change. Against a human baseline on the same specification files, merged agent PRs were breaking less often (9.9% versus 18.9%), contrary to our initial hypothesis; this is exploratory, and the remaining hypotheses are untested.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24296908
Primary Topic
Software Engineering Research
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

When Agents Break Contracts: A Socio-Technical Framework for AI-Authored Pull Requests and API Breaking Changes

Swetha Guruju
Zenodo (CERN European Organization for Nuclear Research)
Software Engineering Research
preprint

When Agents Break Contracts: A Socio-Technical Framework for AI-Authored Pull Requests and API Breaking Changes

Swetha Guruju
preprint en

Abstract

AI coding agents now open pull requests (PRs) on real repositories at scale, and recent mining studies show that many of these PRs are never merged for reasons that go beyond incorrect code: duplicated work, unwanted features, ignored feedback, and reviewers who never engage. In parallel, web API evolution remains a persistent source of client breakage, and mobile clients are especially exposed because users update apps at different times, so old app versions keep calling an API after it has changed. These two literatures have developed independently. This paper synthesizes both and argues that they meet at a concrete, under-studied point: agent-authored changes to API-bearing code are consequential contract-affecting changes that are reviewed by humans who may not have the client-side context needed to judge them. We (i) summarize what is known about failures of agentic PRs and about breaking-change detection and classification, (ii) propose a three-layer socio-technical framework (technical, communication, organizational) extended with a fourth, mobile-specific client-lag layer, (iii) identify five gaps in the current evidence, and (iv) lay out a research agenda with research questions, hypotheses, data sources, and an evaluation protocol for predicting breaking changes before merge. This is primarily a survey and position paper. Its only empirical component is a preliminary analysis of the AIDev corpus: 4,402 of 361,296 agent PRs (1.2%) touch API-definition files, agents differ markedly in how often they do so, and in a 150-PR sample of OpenAPI changes, 14 of the 108 that could be automatically diffed (13.0%) contain at least one breaking change. Against a human baseline on the same specification files, merged agent PRs were breaking less often (9.9% versus 18.9%), contrary to our initial hypothesis; this is exploratory, and the remaining hypotheses are untested.

Zenodo (CERN European Organization for Nuclear Research)
Software Engineering Research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon