Set up a separate sandbox QA pass
Codex turned an existing Claude login into a second reviewer that checked an AI system without changing the user’s credentials.
- What they did
- While auditing defects in their AI system, Codex searched the machine for Claude and found an already authenticated account. It then placed Claude Code in a temporary install location, dealt with a Node runtime setup snag, and started Claude in a no-chat, reviewer-only mode. After that first attempt could not reach the test suite, Codex copied just the audit materials into a fresh isolated directory and ran Claude again with permission to read files and execute commands, while blocking edits and saved sessions.
- How it went
- The post says Claude’s first review largely lined up with the problems Codex had already identified. On the second pass, the sandboxed agent was able to run Python tests and adversarial checks. The main drawback was that the initial run stayed in plan mode and never got to execute tests.
- Tips
- This only worked because Claude was already logged in on the Windows account, and Codex reused that session as-is. Keep the reviewer boxed in, and limit it to the minimum inputs needed for QA.
Try it yourself with Codex
Act as a separate QA agent for [project folder] after I finish [feature or fix]. Work in an isolated environment, inspect [tests to run, edge cases to check, behavior to verify], and report anything that looks broken or risky. Do not use, alter, or copy my credentials or tokens; if access would depend on them, stop and ask me before proceeding. Do not edit code unless I approve a specific change plan. Before any irreversible action such as sending, deleting, submitting, or paying, stop and wait for my approval. Finish with a concise summary of issues found, how you tested, and the next steps.
Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.
Get 3 agent uses like this every morning
With the day's top AI news. Free, and one click to unsubscribe.
Related agent uses
Let an agent turn browser edits into code
A designer changed a local site in the browser while a coding agent wrote the matching CSS updates into the codebase.
Use Claude Code to refactor a monorepo in phases
A developer used Claude Code to swap tooling, reorganize features, and refresh a site by checking each step in a tight validation loop.
Use Claude Code to spot macOS layout bugs
A developer let Claude Code compile a Mac app, open it, and click through screens to find visual problems.
Build internal real estate tools with Codex
A rental owner used Codex to make custom tools for offer pricing, property reports, and contract tracking.
Use a long-running agent to chase a stubborn bug
A developer let Codex work overnight against a repeatable test, using debug prints and hypotheses to track down an unexplained bug.
Use Claude Code from a phone with a remote VM
They accessed Claude Code on an iPhone by pairing a Tailscale-only Vultr VM with Termius, mosh, and tmux.