How OpenAI Codex Works Behind-the-Scenes (and How It Compares to Claude Code) (blog.promptlayer.com)

This deep dive unpacks how OpenAI’s Codex CLI actually operates and contrasts it with Anthropic’s Claude Code. At its core Codex runs a single-agent, ReAct-style master loop (Think → Tool Call → Observe → Repeat) via the AgentLoop.run() function, using OpenAI’s Responses API with streaming and function/tool calls. A powerful system prompt encodes a mini-API (e.g., apply_patch with unified-diff envelopes) and safety/UX rules, nudging the model toward small, auditable diffs and iterative read→edit→test cycles. Tooling is “shell-first”: a general shell executor (cat, grep/find, ls, run tests/git) plus an apply_patch for file mutations. Execution modes (Suggest, Auto-Edit, Full Auto) and OS-level sandboxes (Seatbelt on macOS; Docker+iptables on Linux) enforce approvals and network/filesystem limits. Comparatively, Claude Code uses a structured toolset (View/LS/Glob, GrepTool, Edit/Write, Bash, WebFetch, notebook tools), proactive project scanning, explicit planning (TodoWrite, /think) and finer-grained permission UX. Practical implications: Codex’s minimal, open-source, locally runnable design favors surgical edits, low complexity and strong containment but can hallucinate if files aren’t loaded; Claude’s richer tooling and context-loading excels at large refactors and multi-file reasoning but is more opinionated and verbose. For practitioners, choose Codex for controlled, local iterative workflows and Claude for broad-context, ambitious codebase transformations.