💡 Claude sometimes confidently wraps up an answer on buggy code - and says "done" itself. I used to catch this by eye or when something broke an hour later in a live system. Now before every final answer Codex automatically reads all the edits and gives a verdict
What Codex is and how the logic works:
Codex is an autonomous cloud agent from OpenAI for checking code, announced in May 2025. It runs in an isolated container in read-only mode: it physically can't change anything, only read and analyze. The internet is off while it works - the agent sees only the code and dependencies from the repository
Two AIs, trained by different teams on different data, systematically make mistakes in different places. Claude wrote the code - Codex looks at the same edits independently. What the first one missed, the second one catches. This isn't about who's smarter - it's the second-pair-of-eyes principle that always works in development. At the end Codex writes in Russian: "VERDICT: clean" or "VERDICT: issues found" with a concrete list in the format file:line - the point
How the system is built - three parts:
📌 Auto-run at the finish - when Claude is about to finish its answer, a script fires. It checks three things: are there real edits in the session, did anything actually change, wasn't Codex just run. If everything checks out - it runs the review and returns the analysis before Claude's final answer
📌 Manual run via a skill - at any moment I ask Claude to run the review right now. Handy when I want to rerun it after fixes, or check a piece of code before the work is done
📌 AGENTS.md in every project - a file with instructions for Codex in the repository root: what's allowed and forbidden, style rules, architectural constraints. An open standard - the counterpart of Anthropic's CLAUDE.md. Codex reads the file itself before every review
What Codex looks for in every review:
✅ Bugs and logic errors that are easy to miss when you write fast
✅ Security holes - leftover passwords, unsafe input, SQL injections
✅ Violations of the project patterns spelled out in AGENTS.md
✅ Edge cases the author missed
In March 2026 Codex Security scanned 1.2 million commits of open-source projects and found about 800 critical and over 10,000 high-severity vulnerabilities in Chromium, OpenSSL and PHP. That's the principle in action: even in code that's been reviewed for years - a second independent reviewer finds what everyone before them missed
Available on the ChatGPT Plus subscription $20/month - the base plan, not a separate tier. Codex works in the background in the cloud and returns a result in 1-30 minutes. On the same complex task it spends ~1.5M tokens vs ~6.2M for Claude Code - faster and lighter for a parallel review
💻 Performance: SWE-bench Pro ~57-59% - on par with Claude Code on hard engineering tasks
⭐️ Eric S. Raymond, open-source developer, "The Cathedral and the Bazaar":
"Given enough eyeballs, all bugs are shallow"
A second look isn't only for code
The same principle works for me wherever a mistake is expensive. In the content factory, names, numbers, dates and quotes are checked against the collected material before I even see the draft, and then the post still goes past my own eyes. That has produced 84 posts, and not one went out without my approval. For my personal agent Jes, running for 134 days, the second look is built into the rules: deleting data, spending money, messaging clients and publishing always wait for my yes. I now run 8 Claude Code agents every day, and none of them is allowed to act without a check where the cost of a mistake is high
🎯 It used to be that "a second look at the code" meant waiting for a colleague or rereading it yourself an hour later. Now it's 20 seconds while Codex reads the edits and gives a verdict. $20 a month that saves hours of debugging - the simplest investment in my stack I've made in a long time