Hands-on with seven coding agents tasked with building a website: Claude Code performed best, OpenAI's Codex lacked polish, and Replit failed to build the site
Timothy B. Lee / Understanding AI :
Context & Ripple Effects
The test arrives after coding agents moved beyond code completion toward end-to-end app creation: Replit had positioned its Agent as a tool that could build applications from prompts, while OpenAI introduced Codex CLI for terminal-based local tasks. A common website-building task makes the gap between those product promises easier to compare.
The result matters because it evaluates the finished work rather than only a model’s ability to generate snippets. Claude Code’s lead, Codex’s rougher output, and Replit’s failure show that agentic coding quality remains uneven across products aimed at similar outcomes.
First-order effects
- Claude Code gains third-party evidence of stronger performance on this specific website-building task, while Codex’s output is characterized as less polished.
- Replit’s failure to complete the site under this evaluation directly undercuts its positioning around prompt-to-app creation, at least for this test scenario.
Second-order effects
- Teams comparing coding agents will have more reason to assess end-to-end completion and output quality, not just whether an agent can produce code or invoke tools.
- OpenAI and Replit face pressure to improve the reliability and finishing quality of their agent workflows; Claude’s apparent advantage raises the competitive bar for comparable tasks.
Third-order effects
- If repeatable evaluations keep producing divergent outcomes, coding-agent competition will be shaped less by feature checklists and more by dependable execution across multi-step software tasks.
- The category may split between agents that assist inside established developer workflows and agents trusted to deliver a usable first version of an application; this single test does not establish which model will prevail.
The trend: Coding AI is shifting from autocomplete-style assistance toward competition over autonomous, end-to-end delivery of working software artifacts.