Claude Code vs Codex: Same task, very different blast radii

When software engineers assign a simple coding task to AI coding agents—such as adding input validation to a route handler in a small Node.js/Express service—the task appears straightforward: edit a single file and write roughly 20 lines of code. However, background system monitoring reveals drastically different underlying behaviors between competing tools.

This benchmark test standardized the environment by running Claude Code 2.1.72 and Codex CLI 0.113.0 against a clean copy of a user management service. Every system call was recorded using strace -f to track openat, connect, and execve processes. While both agents successfully completed the task, the digital footprint and operational scope left behind varied significantly.

Stock photo for illustration only, not from the actual event

In terms of file access volume, Claude Code opened 2,779 unique files during the session, whereas Codex opened only 303. Although a large portion of Claude's reads comprised its internal infrastructure such as plugin caches and configuration files, the exact targets of these reads raise important operational security considerations.

During execution, Claude Code attempted to open /proc/<pid>/environ for 752 distinct process IDs, successfully reading 256 of them under the current user. Terminal-launched processes on Linux rely on walking /proc to inherit environment variables exported by parent shells, such as API keys and PATH settings.