scripting languages

研究通过分析GitHub Agentic Workflows的结构和维护方式，探讨了开发者如何定义和维护由AI代理执行的工作流程，并建议增加防御措施。

Early-stage software development—spanning requirements elicitation, testing, and deployment—is hindered by ill-defined tasks and dense manual intervention points, impeding automation. Traditional CI/CD pipelines address only post-coding phases, leaving semantic gaps between underspecified stages unbridged. Method: We propose “workflow-as-software,” a novel paradigm that models end-to-end development as programmable workflows. Leveraging large language models (LLMs) as universal semantic adapters, our approach automatically reconciles heterogeneous task semantics. It integrates domain-specific workflow orchestration, a lightweight domain-specific language (DSL), and semantic translation interfaces. Contribution/Results: Evaluated in production at Volvo, the method reduced test automation effort by 2–3 full-time engineers and compressed the end-to-end development-to-deployment cycle to two months. It marks the first demonstration of LLM-driven, fully automated software delivery across the entire lifecycle—from requirements to deployment—thereby extending automation beyond conventional CI/CD boundaries.

This study addresses the significant burden developers face in authoring and maintaining GitHub Actions workflows, stemming from a lack of systematic understanding of real-world automation and reuse practices. Through a mixed-methods approach combining a survey of 419 practitioners with qualitative and quantitative analysis, this work presents the first developer-centric characterization of common automation tasks, patterns of reuse mechanism adoption, and maintenance pain points in workflow development. The findings reveal that while developers heavily rely on reusable Actions, they seldom adopt reusable workflows; version management challenges lead to rampant copy-pasting; and critical aspects such as security and performance monitoring remain under-automated. These insights provide empirical foundations for improving CI/CD toolchains and reuse mechanisms.

Existing coding agents are largely confined to code generation and lack support for the full workflow lifecycle, including composition, iteration, deployment, and sharing. This work proposes CURATE, a novel system that integrates modular cataloging and FAIR principles into a large language model–driven multi-agent framework to enable human-in-the-loop, end-to-end workflow development and automated execution. Built upon Claude Opus 4.8, CURATE incorporates user-in-the-loop mechanisms and a module registry to facilitate cross-workflow sharing of reusable components. The system successfully reproduces four SeBS-Flow benchmark workflows and automatically constructs a complex anaerobic digestion simulation pipeline, demonstrating its feasibility and effectiveness in supporting comprehensive, collaborative scientific workflow automation.

Non-programmers face significant challenges in creating secure and efficient software script automations, as conventional approaches require programming expertise and API knowledge, while runtime code generation suffers from unverified outputs, security vulnerabilities, high latency, and substantial computational overhead. Method: This paper proposes an offline simulation-driven framework for skill discovery and validation. It treats software script interfaces as system-level testbeds for large language models (LLMs), employs a graph neural network (GNN)-based API coordination prediction model to identify infrequent yet semantically valid API combinations, and integrates top-down functional guidance with bottom-up API coordination exploration—leveraging offline execution feedback for iterative script refinement. Contribution/Results: Evaluated on Adobe Illustrator, the framework achieves markedly higher automation success rates, significantly reduced response latency, and substantially lower token consumption compared to baseline methods.

This study addresses the lack of systematic understanding regarding how GitHub Actions workflows are used in real-world scenarios, how developers respond to workflow failures, and how these practices relate to project characteristics. Combining large-scale quantitative analysis of 258,300 workflow runs with qualitative case studies across 21 diverse repositories, this work identifies three typical patterns developers employ to handle workflow failures and uncovers a “configuration–usage gap”—where YAML configurations exist but workflows remain effectively unused. Furthermore, the study empirically validates five hypotheses linking project features to workflow usage intensity, revealing a significant positive correlation between high usage intensity and low failure rates. These findings provide actionable empirical evidence for improving CI/CD practices.

This study addresses the imprecise evaluation in existing agent skill package self-evolution benchmarks, which fail to distinguish among documentation repair, script repair, and behavior preservation. To overcome this limitation, we construct the first fine-grained skill evolution benchmark comprising 350 tasks that separately evaluates these three capabilities. Furthermore, this work proposes an Abstract Syntax Tree (AST)-guided revision method that leverages static call graph constraints to restrict the editing scope, thereby enabling coordinated updates and precise repairs across both documentation and code. Experimental results demonstrate that the proposed approach achieves an absolute improvement of over 20% in repair success rate compared to a pure Markdown baseline, while significantly enhancing consistency across multiple execution runs.

This study addresses the disconnect between existing coding and computer-use agents, as well as the lack of visual interaction to assist software diagnosis and repair, by being the first to systematically investigate the role of visual feedback in this task. Methodologically, it integrates source-code-level execution, application screenshot analysis, and graphical interaction mechanisms to construct a benchmark environment spanning four domains, requiring agents to extract specification information from runtime interfaces and validate their modifications. The primary contribution lies in providing executable correctness evaluation criteria that systematically quantify the capability of state-of-the-art agents to accomplish software engineering tasks by combining code editing, command execution, and GUI-based visual feedback.

Automatically constructing high-quality, reusable skills from heterogeneous, fragmented interaction traces—often missing critical security behaviors—is highly challenging. This work proposes the W2S framework, which introduces a novel intermediate representation called RWSA to decouple skills into workflow structure, execution semantics, and runtime attachments, thereby enabling task decomposition, control-flow modeling, verification, rollback, and state management. W2S achieves efficient skill construction through trajectory segmentation, local skill draft generation, structural alignment, branch fusion, redundancy compression, and confidence-aware retention. Experimental evaluation across 70 skills demonstrates that W2S improves behavioral replay consistency by 10.5% compared to baseline approaches based on summarization and prompting.

This study addresses the challenge of determining appropriate granularity and responsibility allocation for AI agents in human-AI collaborative software engineering. The authors propose a method that dynamically generates AI agent roles based on project-specific context by integrating object-centric process mining with both imperative and declarative process modeling. Leveraging event logs from software repositories, the approach automatically discovers agent structures aligned with the unique characteristics of a project’s development process and synthesizes corresponding specifications and implementations. Moving beyond predefined role templates, the method has been successfully deployed in prominent open-source projects. Empirical validation through functional testing and user studies demonstrates that the generated agents exhibit responsibility boundaries and collaboration efficiency closely matching developer expectations.