How Developers Configure AI Coding Agents: New Study Maps the Landscape

Agentic AI coding tools are now handling substantial portions of software development workflows, but how developers actually configure these systems remains understudied. A new paper from arXiv examines the configuration mechanisms across five major platforms and reveals emerging patterns in how teams customize their AI assistants.

The Study Setup#

Researchers conducted a systematic analysis covering Claude Code, GitHub Copilot, Cursor, Gemini, and Codex. Rather than focusing on model performance, they investigated the human side of the equation: what repository-level artifacts (Markdown and JSON files) developers use to shape how these tools behave. Their empirical study looked at 2,853 GitHub repositories to determine adoption rates and real-world usage patterns.

Eight Configuration Mechanisms Identified#

The analysis uncovered eight distinct ways developers can steer agentic tools, ranging from static context provision through to executable integrations and external connections. The researchers paid particular attention to three mechanisms: Context Files, Skills, and Subagents.

The findings reveal a clear adoption hierarchy. Context Files dominate the configuration landscape—they're often the only mechanism a repository uses. An emerging interoperable standard, AGENTS.md, is gaining traction across multiple platforms. Meanwhile, more sophisticated mechanisms like Skills and Subagents see far lower adoption rates, suggesting that developers are still learning how to leverage advanced customization options.

The Current Baseline#

The study establishes what appears to be developers' starting point: static instruction files that provide background information without requiring executable code or plugin architecture. Skills, when adopted, tend to rely on plain-language instructions rather than proper executable scripts. Configuration practices are beginning to diverge by tool, with Claude Code users employing the widest variety of configuration strategies.

What's Next#

These findings create a baseline for understanding the evolving relationship between developers and AI agents. The authors argue that AGENTS.md should serve as a natural foundation, and call for future research on how configuration approaches change over time and affect agent performance. .