Claude Code, Codex, and Pi Agent Loops Compared: Where They Agree and Diverge

A technical deep dive published in September 2026 compared the agent loop architectures of three AI coding tools — Anthropic's Claude Code, OpenAI's Codex, and Pi — across codebases ranging from 794 to 1,729 lines. All three independently converge on the same five-step control flow: assemble context, call the model, execute tool calls, append results, and decide whether to loop again. Despite this shared pattern, the teams diverge significantly in how they handle edge cases such as streaming responses, parallel tool execution, mid-turn interruptions, and provider failures. The analysis also corrects a previously published claim that open-source agent loops run in roughly 50–200 lines, noting that even the most minimal production harness examined — Pi — requires 794 lines. The author attributes the gap not to open-source versus proprietary code, but to the fundamental difference between demo implementations and tools built for sustained real-world use.