An AI coding agent is a model wrapped in a harness that can read files, run commands and check its own work, rather than a chat window that returns a code block. The two most used agents in October 2026 are Anthropic's Claude Code and OpenAI's Codex, and every build on this site was made by one of them. That makes this shelf a direct comparison: the same eight briefs, each agent running its vendor's top models, one attempt, nothing edited.
What the harness changes, in practice:
- How long the agent keeps going. Claude Code runs took from 25 minutes to more than two hours, and one hit our 150-minute cap. Every Codex run finished in 7 to 25 minutes, each in a single turn.
- Whether it tests itself. Both agents may open their own page in a browser to check it. The Claude runs used that freedom far more, which shows up as many turns and large input-token counts.
- What "best coding agent" means. If you want a finished artefact in ten minutes, the answer differs from the one you get if you want the most considered result and can wait an hour.
The pages below compare the agents head to head, review each one by its sixteen builds, and show what each was asked and what it returned. The exact command lines are on the methodology page.