Loop Engineering for Agentic Software Systems: LEAF Architecture, Pattern Taxonomy, Use Cases, Evaluation, Challenges and Future Directions

AI coding agents are shifting software development from isolated prompt–response interaction towards iterative, tool‐mediated, stateful, and feedback‐driven workflows. This emerging practice, increasingly described as loop engineering , moves the unit of design from the individual prompt to the recurring agentic process that plans, acts, observes, verifies, remembers, and continues under defined governance conditions. Although related work on coding agents, context engineering, tool use, human‐agent interaction, and agent harness engineering has examined important parts of this transition, the loop itself remains underdeveloped as a first‐class software‐engineering design object. This article conceptualises loop engineering as a systems‐level paradigm for agentic software systems. It distinguishes loop engineering from prompt engineering, context engineering, harness engineering, workflow automation, DevOps automation, and general agentic AI. It then proposes LEAF , a Loop Engineering Architecture Framework organised around control, execution, and assurance planes. LEAF models loop‐engineered systems in terms of bounded goals, triggers, context construction, agent planning, tool‐mediated action, observation, verification, memory, governance, escalation, and auditability. The article further develops a taxonomy of loop patterns, including retry, plan–execute–verify, explore–narrow, test‐driven, compiler‐driven, review‐driven, runtime‐debugging, human‐in‐the‐loop, maker–checker, and continuous‐maintenance loops. Possible applications are discussed across debugging, test repair, dependency migration, CI triage, pull‐request review, documentation maintenance, security scanning, DevOps monitoring, educational programming support, and research automation. The article also analyses major risks, including weak verification, context drift, cost escalation, unsafe autonomy, test overfitting, comprehension debt, cognitive surrender, benchmarking difficulty, multi‐agent coordination, and unclear accountability. Finally, it outlines a research agenda covering formal loop specification, trustworthy verifier agents, cost‐aware scheduling, loop observability, benchmark design, permission models, multi‐agent coordination, and human‐centred governance.