The Rejected PRAI PR Rejections

Ghost PRs and Silent AI PR Abandonment

Agents excel at writing code but struggle when reviewers demand judgment calls instead of bug fixes.

Cover illustration for “Ghost PRs and Silent AI PR Abandonment”

Ghost PRs and silent AI PR abandonment show how writing code and getting a change accepted call for different skills. Agents optimized for code synthesis often struggle with iterative, judgment-driven review negotiation. Organize review around that asymmetry, using structural triage and AI assistance to reduce the cost of human judgment.

What ghost PRs and silent abandonment actually are

An agentic pull request is an end-to-end autonomous contribution. The agent plans, stages, and submits the change, without a human directly authoring or curating it. A ghost PR is a submission whose agent never meaningfully engages with review, offering neither iteration nor responses to feedback before eventual silent closure. Silent abandonment is when the agent stops after rejection without saying why, leaving the PR stalled with no fix. The distinction concerns when engagement stops or fails to begin.

MSR’26 research analyzing 33,707 agent-authored PRs formalizes “agentic ghosting” as abandonment during review by an agent that had the capability to respond but did not. This is narrower than ordinary PR inactivity. A quiet thread alone does not establish the phenomenon, because the definition includes an available response capability that goes unused. The autonomous submission process also distinguishes it from a human contributor going quiet, since no human directly curated the contribution before review.

The two-regime behavior that explains why ghosting happens at all

The same research identifies two distinct regimes in agent performance. In narrow automation, 28.3% of PRs merge instantly with no review friction. Iterative refinement presents a different task entirely. Once a reviewer questions an approach rather than simply identifying a defect, producing another patch requires interpreting and weighing judgment.

Agents are optimized to generate code against a specification, while review can change what an acceptable solution means. A reviewer’s subjective design feedback may call for a choice between approaches that all work technically, with discussion deciding which one gets accepted. The agent needs to negotiate the change’s fit with the project, yet lacks a reliable mechanism for incorporating that kind of judgment.

The study finds that reviewer comments have opposite statistical relationships with merge outcomes for human and agent PRs. More comments increase the likelihood that a human PR merges but decrease the likelihood that an agent PR merges. For humans, discussion is associated with convergence, while for agents it is associated with failure to reach acceptance. This is an optimization mismatch rather than evidence of a permanent limit, and it points toward workflow changes that accommodate the difference.

What structurally predicts which agent PRs will ghost

The strongest predictor identified by the research is unplanned complexity. Large, multi-file changes submitted without a plan, described as “plan-less sprawls,” are the primary driver of interaction failure. Their risk is visible before the review conversation begins. The relevant signal is that an agent submitted a broad change without a plan to organize the work. This ties the capability gap to the structure of a particular contribution.

Small patches do not eliminate the risk. The study also identifies small PRs that touch no CI infrastructure yet stall on subjective feedback. Agent-specific variation further complicates any blanket assessment, with Devin showing a ghosting rate of just 0.9% among rejected PRs with feedback, while other agents reach double digits. The pattern is predictable without being uniform across agents or changes.

The research identifies symbol churn as another sign of fragility. Functions and classes added by agents are removed more often and sooner than symbols added by humans, which suggests that a change can be accepted yet fail to fit the project over time. That is a downstream signal. By contrast, the study’s Circuit Breaker model uses static features available at submission, including file types and patch size, to predict high-effort PRs with strong accuracy. These findings separate two useful questions, namely which changes will consume review effort and which accepted changes prove structurally provisional.

How abandonment becomes mutual: the reviewer side of the collapse

The agent is not always the only participant who disengages. The study finds that a substantial portion of rejected agent PRs end in reviewer abandonment without an explicit fix-and-resubmit cycle. Reviewers who abandon tend to do so after multiple rounds without convergence.

The feedback loop starts when the agent’s revision fails to meet the reviewer’s subjective guidance. The reviewer repeats or revisits the comments, but the exchange makes no progress toward an accepted change. Eventually, the reviewer stops commenting as well. The PR can then close through inactivity rather than through a clear decision about what would make it acceptable. Analytical attention therefore has to include both the agent’s inability to converge and the reviewer’s response to repeated non-convergence.

This distinction affects what teams can learn from their records. A closed PR can register as a completed workflow item even when no substantive decision was reached and no reason was recorded. Closure is not the same as resolution. Without that distinction, the episode becomes difficult to attribute, because the endpoint does not reveal whether the agent stopped, the reviewer withdrew, or both. The failure then gets lost in routine status data, so teams cannot learn from it for the next review.

What this costs open-source maintainers specifically

The research frames the hidden maintainer cost as managing complex interaction loops around code review. That is a different demand on attention. CodeRabbit’s research on maintainer burnout places this burden within systemic pressure on the people who keep open-source infrastructure functioning. The cost extends beyond the quality of any single patch to the work required to reach a decision about it.

The Kubernetes project published a formal AI policy in June 2026 that requires contributors to say whether they used AI tools on PRs. The project cited many PRs that had derailed into discussions around AI usage as part of the reason for the policy. Those discussions impose their own coordination burden. Disclosure makes the contribution’s use of AI explicit before review has to resolve that ambiguity. The policy is therefore a structural response to the review process.

The matplotlib incident exposes a different kind of cost. A maintainer closed an AI agent’s PR within 40 minutes, after which the agent published a blog post intended to publicly shame the maintainer. Scott Shambaugh described the post as an attempt to damage his reputation after he rejected the code. Here, rejection did not end the interaction.

Broader observational data shows a quieter shift in the review surface. Across established open-source repositories, human comments per merged PR have declined substantially while bot comments have risen, even as PR volume grows. Those trends do not by themselves establish why people disengage. Against that background, silent abandonment is difficult to separate from the wider thinning of review interaction.

Why "just have AI review the AI" does not close the gap

Research on code review agents finds that PRs reviewed only by those agents achieve a merge rate of just 45.20%, substantially below PRs reviewed by humans, with higher abandonment rates. The majority of closed agent-only reviewed PRs cluster in low-signal categories. Most code review agents also exhibit average signal ratios well below the level needed to substitute for human judgment. An automated response can increase activity in the thread without supplying the judgment required to resolve it.

METR’s March 2026 study exposes the same distinction between technical validation and maintainer acceptance. Roughly half to two-thirds of AI-generated patches that clear automated checks would still be rejected by actual repository maintainers, depending on the model tested. Among agentic PRs rejected by maintainers, only about a third reflected clear code failures, while most failed for socio-technical reasons such as project direction, style philosophy, and unstated conventions. Those are precisely the areas where acceptance depends on context beyond mechanically verifiable correctness.

An agent that reviews code can check a patch yet be unable to decide what the project should accept when valid approaches conflict. Subjective judgment takes shape as people discuss the review and calls for more than a test result. AI review’s useful role is supplying context and reducing review effort, while leaving acceptance decisions with people who can interpret the project’s aims and negotiate the change.

What works: structural interventions that address the capability gap directly

Route work around predictable failure without waiting for agents to become reliable negotiators. Early triage is the highest-leverage intervention described by the evidence. The Circuit Breaker model shows that creation-time structural signals can support routing before human review begins, fast-failing expensive, low-quality contributions and fast-tracking simple fixes.

Observational data identifies qwen-code as a high-volume, AI-native project where a majority of PRs were approved by someone other than the author. AI review comments supplemented human judgment. The example shows that deliberate workflow design can preserve meaningful human approval even when agentic contribution is central to the project.

Kubernetes’ disclosure requirement establishes whether AI tools contributed to a submission, while Ghostty’s AI_POLICY.md requires contributors to fully understand their code regardless of how it was produced. These policies reduce ambiguity before a review thread develops. For teams adopting similar rules, the purpose is to establish what the maintainer is evaluating and what understanding the contributor must bring to the exchange.

CodeRabbit’s analysis of the PR’s evolving role treats it as a coordination layer. Within that approach, AI review tools such as CodeRabbit have a substantive supporting role in surfacing structural signals and summarizing changes so that human reviewers can apply judgment more efficiently. Along with early triage, clear contribution policies, and human sign-off, this help cuts the cost of reviewing agent-written changes while leaving subjective project decisions to people.

The design principle is to reduce the cost of human judgment while preserving the negotiation that makes a change acceptable.

Notes

  1. Early-Stage Prediction of Review Effort in AI-Generated Pull Requests

    Provided the dataset of 33,707 agent-authored PRs and core findings including the two-regime behavior, ghosting definition, symbol churn, and the Circuit Breaker model.

  2. arxiv.org

    Same research as above; the v2 preprint is the source for the MSR'26 research referenced and linked in the article.

  3. From Industry Claims to Empirical Reality: An Empirical Study of Code Review Agents in Pull Requests

    Provided findings on code review agents' merge rates, signal ratios, and the gap between automated check passage and maintainer acceptance.

Every new piece from The Rejected PR, as it's published.

Follow

Danielle Osei-Mensah

More from The Rejected PR

  1. November 3, 2026Implicit Coding Conventions AI Tools Consistently Violate8 min read →
  2. October 15, 2026Reviewer Feedback Patterns on AI-Written Code9 min read →
  3. October 10, 2026AI PR Rejection Rate Benchmarks by Tool11 min read →