arXiv:2510.16786v2 Announce Type: replace-cross 
Abstract: LLM-powered coding agents, which operate in iterative loops (turns) to solve software engineering tasks, are becoming increasingly powerful. However, their practical deployment is hindered by significant and unpredictable costs. This challenge arises from a combination of factors: quadratically growing token counts with each turn, the high price of models, the large number of turns required for real-world tasks, and the tendency of agents to take inefficient or unnecessary actions. While existing research focuses on optimizing individual turns, the strategic control of the total number of turns remains an underexplored area for managing agent performance and cost. To address this gap, we conduct a comprehensive empirical study on SWE-bench using three state-of-the-art models and evaluate the impact of three distinct turn-control strategies: an unrestricted baseline, a fixed-turn limit with reminders, and a novel dynamic-turn strategy that grants extensions on-demand. Our findings first reveal a fundamental trade-off in the unrestricted setting, where no single model excels across performance, cost, and turn efficiency. We then show that a fixed-turn limit, specifically at the 75th percentile of the baseline, serves as a "sweet spot", substantially reducing costs (by 24%-68%) with minimal impact on solve rates. Most significantly, the dynamic-turn strategy consistently outperforms fixed-limit approaches, achieving comparable or better solve rates while further reducing costs by an additional 12%-24% by intelligently allocating resources only to tasks that need them. This work provides the first systematic analysis of turn-control strategies, offering simple yet effective guidelines for developers to balance cost and efficacy. We demonstrate that dynamic resource allocation is a superior, easy-to-implement approach for deploying powerful yet economically viable coding agents.

أظهرت دراسة تجريبية حديثة استراتيجيات التحكم في الأدوار لوكلاء الترميز المدعومين بـ LLM، الذين يواجهون تحديات بسبب ارتفاع التكاليف وعدم الكفاءة خلال المهام الهندسية البرمجية التكرارية. تقيم الدراسة ثلاث استراتيجيات على SWE-bench، بما في ذلك حد ثابت للأدوار واستراتيجية ديناميكية، بهدف تحسين عدد الأدوار التي يتخذها هؤلاء الوكلاء.

Un estudio empírico reciente ha examinado las estrategias de control de turnos para agentes de codificación impulsados por LLM, que enfrentan desafíos debido al aumento de costos e ineficiencias durante las tareas de ingeniería de software iterativas. El estudio evalúa tres estrategias en SWE-bench, incluyendo un límite fijo de turnos y una estrategia de turnos dinámica, con el objetivo de optimizar el número de turnos tomados por estos agentes.

Une étude empirique récente a examiné les stratégies de contrôle des tours pour les agents de codage alimentés par LLM, qui rencontrent des défis en raison de l'augmentation des coûts et des inefficacités lors des tâches d'ingénierie logicielle itératives. L'étude évalue trois stratégies sur SWE-bench, y compris une limite fixe de tours et une stratégie dynamique, visant à optimiser le nombre de tours pris par ces agents.

A recent empirical study has examined turn-control strategies for LLM-powered coding agents, which face challenges due to rising costs and inefficiencies during iterative software engineering tasks. The study evaluates three strategies on SWE-bench, including a fixed-turn limit and a dynamic-turn strategy, aiming to optimize the number of turns taken by these agents.

More with Less: An Empirical Study of Turn-Control Strategies for Efficient Coding Agents

arXiv:2512.02393v1 Announce Type: cross 
Abstract: Agentic systems are modern software systems: they consist of orchestrated modules, expose interfaces, and are deployed in software pipelines. Unlike conventional programs, their execution (i.e., trajectories) is inherently stochastic and adaptive to the problem they are solving. Evaluation of such systems is often outcome-centric, judging their performance based on success or failure at the final step. This narrow focus overlooks detailed insights about such systems, failing to explain how agents reason, plan, act, or change their strategies over time. Inspired by the structured representation of conventional software systems as graphs, we introduce Graphectory to systematically encode the temporal and semantic relations in such software systems. Graphectory facilitates the design of process-centric metrics and analyses to assess the quality of agentic workflows independent of final success.
  Using Graphectory, we analyze 4000 trajectories of two dominant agentic programming workflows, namely SWE-agent and OpenHands, with a combination of four backbone Large Language Models (LLMs), attempting to resolve SWE-bench Verified issues. Our fully automated analyses reveal that: (1) agents using richer prompts or stronger LLMs exhibit more complex Graphectory, reflecting deeper exploration, broader context gathering, and more thorough validation before patch submission; (2) agents' problem-solving strategies vary with both problem difficulty and the underlying LLM -- for resolved issues, the strategies often follow coherent localization-patching-validation steps, while unresolved ones exhibit chaotic, repetitive, or backtracking behaviors; (3) even when successful, agentic programming systems often display inefficient processes, leading to unnecessarily prolonged trajectories.

قدمت دراسة حديثة إطار عمل يسمى Graphectory، مصممًا للتحليل المتمركز حول العمليات للأنظمة البرمجية الوكيلة، التي تتميز بتنفيذها العشوائي والتكيفي. يهدف هذا النهج إلى تقديم رؤى أعمق حول كيفية عمل هذه الأنظمة بعيدًا عن تقييم النتائج، مع التركيز على تفكيرها وتخطيطها وتكيفاتها الاستراتيجية على مر الزمن.

Un estudio reciente presentó Graphectory, un marco diseñado para el análisis centrado en procesos de sistemas de software agenticos, que se caracterizan por su ejecución estocástica y adaptativa. Este enfoque tiene como objetivo proporcionar una comprensión más profunda de cómo operan estos sistemas más allá de la evaluación de resultados, centrándose en su razonamiento, planificación y adaptaciones estratégicas a lo largo del tiempo.

Une étude récente a introduit Graphectory, un cadre conçu pour l'analyse centrée sur les processus des systèmes logiciels agentiques, qui se caractérisent par leur exécution stochastique et adaptative. Cette approche vise à fournir des informations plus approfondies sur le fonctionnement de ces systèmes au-delà de l'évaluation des résultats, en se concentrant sur leur raisonnement, leur planification et leurs adaptations stratégiques au fil du temps.

A recent study introduced Graphectory, a framework designed for the process-centric analysis of agentic software systems, which are characterized by their stochastic and adaptive execution. This approach aims to provide deeper insights into how these systems operate beyond mere outcome evaluation, focusing on their reasoning, planning, and strategic adaptations over time.

More with Less: An Empirical Study of Turn-Control Strategies for Efficient Coding Agents

Was this article worth reading? Share it

NootCode

Swapcode AI

AQ