Codex inside T3 Code: the 4-minute job that took 10
Here is a number that surprised me: the same Codex model, the same task, the same commit, took 10 minutes 26 seconds inside T3 Code and 6 minutes 07 seconds through the native CLI. Before I blamed the orchestrator, I wan
Here is a number that surprised me: the same Codex model, the same task, the same commit, took 10 minutes 26 seconds inside T3 Code and 6 minutes 07 seconds through the native CLI. Before I blamed the orchestrator, I wanted to know which of those minutes were the model's.
The answer is that 6 minutes 6 seconds of the T3 run were spent waiting for a human to approve one cleanup step.
The question behind the numbers
People who run Codex through an orchestrator often report that it "got slower and dumber." Those are two different claims. Slower is a scheduling question and can be answered from timestamps. Dumber is a quality question and needs the output itself.
I tackled the scheduling question first, then the quality one.
The setup I used
Three runs on commit 3974f307, each in its own git worktree, all building the same single-file offline article-topic tool.
- Native
codex execwith gpt-6-sol. - Codex inside T3 Code with gpt-6-sol.
- Claude Sonnet 5.5 inside T3 Code, as a reference point only.
Only the first two share a model, so they are the one clean comparison. The sandbox also differed: the native run used workspace-write, and the T3 run used danger-full-access, which is wider. The T3 run did not stall because its sandbox was tight. It stalled because T3 asks before any action that leaves the workspace. Every call went through https://api.crazyrouter.com/v1 as an OpenAI-compatible endpoint.
Following the clock
The T3 Codex run finished its feature and self-test at 03:01:24. Then it tried to remove a Chrome temporary profile it had created itself. That path is outside the workspace, so T3 stopped and asked.
I read the approval wait to the moment the model resumed. It took 6 minutes 6 seconds, and 44 seconds of that came after I had already decided.
Strip the wait out and the T3 run's working segment is 4 minutes 20 seconds, 1 minute 47 seconds faster than the native run.
Why the native run was slower in the first place
The native run made 20 tool calls against 7 for T3. Most of the extra calls were checks on the environment: git status, locating the working directory, hunting for an AGENTS.md, testing whether playwright, puppeteer, and jsdom were installed, and checking node and WebSocket support.
The orchestrator had prepared that environment and drawn the boundaries in advance, so the T3 run never needed those checks.
The native run also tried to delete the same temporary profile first and got blocked by policy. The refusal came back at once. The agent switched to fs.rmSync and passed in about three seconds. A refusal that returns in milliseconds is a different experience from one that waits for a person.
The browser cost the native run about 2 minutes 24 seconds. Edge created a profile without opening its debug port, computer-use could not reach the window, and the run ended with a DOM mock under Node and vm.
Quality
I checked the three builds against the same eight requirements, and I counted a hard failure only for unauthorized edits or invented files. None of the builds failed that way.
The native build set a 680-pixel breakpoint but never viewed it in a browser, and its report says so. The T3 Codex build was tested at 360 pixels with no horizontal scroll. The Claude build passed 28 checks in Edge.
Two defects, two of three builds
A Claude thread reviewed the Codex output without knowing its author. It found two reproducible bugs.
Cross-tab writes get overwritten, because the page ignores the storage event and keeps state in memory. Keyboard focus drops to <body> after each action, because the list is re-rendered in full.
The Claude build has the focus bug but avoided the overwrite, since it wrote its own tab sync. The native build has the overwrite bug, because it has no storage listener at all.
I read this as a property of the task. A list with add, delete, and full re-render makes the simplest implementation fail, whoever writes it.
What I'd do differently
If an orchestrated run looks slow, look at the approval queue before the model. Approving safe cleanup ahead of time removes most of the wait. Writing the allowed paths into the prompt cuts down the stops. And a single review pass at the end, after parallel work, caught both bugs in 41 seconds.
What this does not show
This is one run per arm. A 1 minute 47 second difference can come from network or provider queueing, so a real claim needs three runs per arm at least. I did not try a shorter AGENTS.md, so the cost of prompt length is still unmeasured.
๐ Try it with a single API key: https://crazyrouter.com/en/pricing?utm_source=devto&utm_medium=article&utm_campaign=offsite_article&utm_content=cr_d87e0066e7b4_bottom_cta
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ full credit and traffic to the original publisher.