Loop Engineering: Building Blocks, Adoption, and Impact (Supplementary Material)

Over the past months, the way developers direct agentic AI coding tools has moved up several levels of abstraction: from phrasing individual prompts, to engineering the context that agents see, to configuring the harness that surrounds the model. In June 2026, practitioners began to describe a new abstraction layer called loop engineering. Instead of prompting an agent interactively, developers now design systems that prompt agents for them. They specify what to build up front, e.g., using machine-checkable goal conditions, or the goals are derived from repository events such as failing builds and open issues. The term loop engineering spread rapidly through blog posts, newsletters, and social media, accompanied by bold claims and equally loud skepticism, but it has not yet been studied scientifically. In this paper, we trace the trajectory from prompt engineering to loop engineering and report an exploratory review of the emerging gray literature, situated against adjacent academic work. The practitioner sources largely agree on what a well-engineered loop contains: scheduled or goal-driven agent runs, persistent state files, verifier sub-agents, token and cost budgets, and defined points of escalation to humans. From this review, we derive a research agenda centered on the empirical study of loop engineering in open-source projects. We analyze which aspects of loop engineering are traceable from repository data and which are not, and we report an exploratory mining study of 36,710 open-source repositories classified as engineered software projects. Autonomous agent loops run in open-source projects: 70 of the 81 candidate repositories we manually verified execute scheduled or event-triggered agent processes. Yet almost none commit the state files, budgets, and run logs the discourse prescribes, because these loops keep their state in platform artifacts such as labels, issues, and pull requests rather than under version control. As the planned journal extension, we outline a comparative study of agent autonomy levels, together with re-running and scaling this mining study to track how the practice evolves.