Understanding the Architecture of Coding Agents: An Exploratory Study Using a Research Prototype

A systematic architectural description of coding agents, the Ark research prototype, and the ArkBench benchmark

Marco Tulio Valente· 2026· DOI 10.48550/arXiv.2608.10934

The core problem

Coding agents have rapidly emerged as the primary interface for AI-assisted software development. Despite their growing adoption, relatively little is known about their internal architecture, and no systematic architectural description comparable to those available for compilers or operating systems currently exists. This paper addresses this gap by documenting the main architectural components of coding agents, explaining their responsibilities, interactions, and execution flow. To support this effort, the authors present Ark (Agent Research Kit), a minimal open-source coding agent designed for research and education that preserves the essential architectural mechanisms of modern coding agents while emphasizing simplicity and clarity. They also introduce ArkBench, a lightweight benchmark comprising ten representative software maintenance and evolution tasks. Using gpt-5.4-mini, Ark successfully solved 8 of the 10 tasks while requiring modest token consumption. Finally, the architecture of Ark is compared with those of state-of-the-art coding agents using a recently proposed architectural taxonomy. The authors hope that both Ark and ArkBench provide a practical foundation for teach

Using gpt-5.4-mini, Ark successfully solved 8 of the 10 tasks in ArkBench while requiring modest token consumption. The benchmark comprises ten representative software maintenance and evolution tasks, providing a lightweight but practical evaluation suite. The high success rate (8/10) indicates that a minimal coding agent that preserves essential architectural mechanisms can effectively handle realistic software maintenance and evolution tasks. The modest token consumption suggests that the architecture is efficient and does not rely on excessive model calls or context. These results support the claim that Ark is a viable foundation for teaching, research, and experimentation on coding agents. The paper does not report per-task results, timing, or comparisons against other agents on ArkBench; the primary quantitative outcome is the 8/10 success rate with modest token usage.

Coding agents have rapidly emerged as the primary interface for AI-assisted software development. Despite their growing adoption, relatively little is known about their internal architecture, and no systematic architectural description comparable to those available for compilers or operating systems currently exists. This paper addresses this gap by documenting the main architectural components of coding agents, explaining their responsibilities, interactions, and execution flow. To support this effort, the authors present Ark (Agent Research Kit), a minimal open-source coding agent designed for research and education that preserves the essential architectural mechanisms of modern coding agents while emphasizing simplicity and clarity. They also introduce ArkBench, a lightweight benchmark comprising ten representative software maintenance and evolution tasks. Using gpt-5.4-mini, Ark successfully solved 8 of the 10 tasks while requiring modest token consumption. Finally, the architecture of Ark is compared with those of state-of-the-art coding agents using a recently proposed architectural taxonomy. The authors hope that both Ark and ArkBench provide a practical foundation for teaching, research, and experimentation on coding agents.

The study follows an exploratory design aimed at producing a systematic architectural description of coding agents. The authors first identify the essential architectural mechanisms common to modern coding agents and formalize them into a reference architecture. They then implement Ark (Agent Research Kit), a minimal open-source coding agent that preserves these mechanisms while prioritizing simplicity and clarity for research and education. Ark is evaluated using ArkBench, a lightweight benchmark of ten representative software maintenance and evolution tasks. The evaluation uses gpt-5.4-mini as the underlying language model and measures task success and token consumption. Finally, the architecture of Ark is compared with those of state-of-the-art coding agents using a recently proposed architectural taxonomy. This methodology combines architectural documentation, prototype implementation, empirical benchmarking, and comparative analysis to bridge the gap between ad hoc descriptions and systematic architectural knowledge.

The paper contributes a systematic architectural description of coding agents, addressing the absence of documentation comparable to that of compilers or operating systems. Ark serves as a concrete, minimal reference implementation that preserves essential architectural mechanisms while remaining simple and clear, making it suitable for research and education. ArkBench provides a lightweight benchmark of ten representative software maintenance and evolution tasks, enabling reproducible experimentation. The comparison of Ark with state-of-the-art coding agents using a recently proposed architectural taxonomy situates the prototype within the broader design space and highlights architectural commonalities and differences. The 8/10 success rate with modest token consumption demonstrates that a minimal architecture can be effective for realistic tasks. The authors position Ark and ArkBench as a practical foundation for teaching, research, and experimentation on coding agents. Limitations include the exploratory scope, the single model used (gpt-5.4-mini), and the small benchmark size; future work may extend the taxonomy, expand ArkBench, and evaluate additional models and architectural variants.

CS practitioners and researchers

Opening member content…