I Told the AI “Don’t Stop,” and It Still Stopped — OpenAI Officially Documents GPT-6 Astra’s Approval-Pause and Over-Testing Habits
1. It started with: “Is this actually official?”
A social post claimed that OpenAI had published guidance for making GPT-6 Astra work more autonomously.
The advice was suspiciously specific: infer intent and scope from context, continue through authorized work instead of requesting approval at every step, do read-only work, reviews, and reversible fixes without repeatedly asking permission, and move any genuinely necessary approval to the last step after producing something concrete and reviewable.
It also claimed that Astra can be unusually sensitive to instructions in files such as SKILL.md and AGENTS.md, and that on small coding changes it may test more broadly than the task actually requires.
That sounded less like generic documentation and more like a list of the exact behaviors people complain about.
So the only sensible move was to check OpenAI’s own documentation.
It was there.
Almost exactly as advertised.
2. OpenAI literally names “unnecessary approval pauses”
OpenAI’s GPT-6 Astra model guidance says the model is more likely than earlier models to ask a question when additional input could materially change the result.
That is useful collaboration behavior.
It is also how you get this:
User: “Investigate and fix everything.”
AI: “I finished investigating. May I apply the fixes?”
User: “I literally told you not to stop.”
AI: “Understood. Here is the fix plan. Should I proceed?”
Congratulations. We have invented a permission-powered perpetual-motion machine.
OpenAI’s migration quickstart even includes the phrase Unnecessary approval pauses. This is not just a user imagining a pattern. The vendor has a name for it.
It is like walking into a clinic and hearing, “Yes, that is excessive-confirmation syndrome. Next.”
3. Why “don’t stop” was too vague
“Don’t stop” states a goal, but it does not define the model’s decision boundary.
The default collaboration instinct remains: if one more question could improve the result, ask it.
OpenAI’s recommended prompting is more operational. Infer the user’s intent and task scope from instructions and prior conversation. Treat action-oriented requests as instructions to do the work. Do not stop at “yes, I can,” a plan, or a partial result chosen merely to save time, effort, or tokens. Do not repeatedly seek approval for read-only work, reviews, reversible fixes, or work already authorized or strongly implied by the task.
So:
“Don’t stop” is a motivational poster. The official guide is a map of where stopping is actually forbidden.
No surprise which one works better.
4. Approval should be the final red button, not an entry ticket
The most useful design idea in the official guidance is to move approval to the end.
Suppose a site needs to be fixed and then published, and only the final publish action genuinely requires confirmation.
The agent should still investigate, edit, build, run the necessary tests, and prepare a reviewable diff first. Then it can ask one concrete question: “Everything is ready; only publishing remains.”
That is very different from:
“May I investigate?”
“May I edit?”
“May I test?”
“May I test again?”
Users asked for work, not an administrative stamp-collecting minigame.
5. OpenAI also documents the “testing too much” problem
The second official behavior description is just as recognizable.
OpenAI says Astra tends to be thorough about testing coding changes before considering a task complete. For smaller tasks, that can result in broader testing than the task requires.
You can picture the sequence.
Change one line.
Run the relevant test.
It passes.
Add another test “just in case.”
It passes.
Add a low-value regression test that mostly mirrors the implementation.
It passes.
Meanwhile, the actual task still has not shipped.
Testing has stopped being quality control and become a hobby.
OpenAI recommends calibrating verification to the change: do not write tests for reversible, low-impact changes when they merely mirror the implementation; run meaningful tests and required checks; once they pass, broaden or repeat testing only when new changes, failures, or unresolved concerns justify it.
Tests are not the destination. They are pre-shipment inspection.
If the product passed inspection, do not send it back to the warehouse just so you can invent another checklist.
6. “Then make that the default” is a pretty natural reaction
If OpenAI understands the behavior this clearly, the obvious reaction is: why not make stronger follow-through the default?
There is a legitimate tradeoff. Astra is also designed to respect task boundaries and ask focused questions when missing information could change the outcome. A general-purpose assistant that interprets “fix my site” as permission to publish anything it wants would create a different class of problem.
The frustration begins when the user has already made the boundary clear — finish the authorized work, continue through reversible steps — and the model still falls back to caution at every junction.
That is why the official guide tells developers how to prompt for more autonomous execution.
A safe default makes sense.
But power users will still think:
“Great. Now put the Autonomous switch in the dashboard instead of hiding it in the owner’s manual.”
7. Codex keeps the full global autonomy prompt
For Codex, a two-line rule is not enough. It may need to read a repository, modify files, resolve conflicts, test code, prepare a PR, and work for a long time without turning every stage into a permission checkpoint.
So the full all-session Codex prompt stays intact.
It begins by requiring current OpenAI documentation to be checked and the real environment to be read: $CODEX_HOME, config.toml, AGENTS.md, skills, and global instructions. It treats global configuration such as developer_instructions as a canonical candidate while preserving existing settings and keeping the change idempotent.
It then defines intent inference, no early stopping at plans, already-authorized reversible work, approval-last, clarification-last, instruction conflicts, transparency about the exact SKILL.md that caused a pause, no invented approval bureaucracy from hypothetical risk, proper boundaries for destructive or irreversible operations, useful subagent parallelism, proportionate testing, surface readback, smoke tests, idempotency, and explicit completion criteria.
In other words, Codex does not merely get “keep going.” It gets:
“Here is the whole traffic code, here are the only red lights, and if you stop somewhere else, show me the exact file that told you to.”
That stays.
8. ChatGPT only needed two persistent rules
Daily ChatGPT use did not need the entire Codex policy.
Markdown is not a problem.
Lists are not a problem.
Fine-grained subagent policy is not a daily irritation.
The repeated pain points were only two:
Stopping for unnecessary confirmation when the request and scope are already clear.
Expanding tests beyond what is necessary and meaningful.
So the ChatGPT-side persistent rule was reduced to exactly those two ideas.
When intent and scope are sufficiently clear, do not end authorized work with “Would you like me to continue?” unless a real higher-level safety, permission, or irreversible-action boundary requires it.
Run tests appropriate to the change. Once the necessary checks pass, do not keep repeating or multiplying tests unless new changes, failures, or unresolved concerns give a reason.
Fix the things that are actually annoying. Do not turn the policy itself into new noise.
9. This is closer to “reduce stop-loss” than “save Astra quota”
The social post framed these recommendations as a way to save Astra usage.
That intuition is understandable, but OpenAI presents them as Prompting best practices for tuning model behavior. The documentation does not promise that adding these instructions reduces usage by a fixed percentage.
In fact, an agent that truly follows through may work longer in one run because it actually finishes the job.
What you can remove is the dead time and repeated context caused by:
ask → wait → resume → ask again.
So the better description is:
Spend less of the run on permission theater and more of it on completion.
10. Conclusion: keep moving, inspect enough, finish the job
After digging through cutting-edge AI guidance, the final operating rule is hilariously ordinary.
Stop looking back at the human after every step.
If inspection passed, ship the thing.
Codex gets the full traffic code. ChatGPT gets those two signs on the wall.
And the funniest part is that this is not merely a homemade “AI training trick.” Much of the diagnosis and the recommended correction is explicitly described in OpenAI’s own guidance.
We went looking for the future of AI and ended up at a factory supervisor’s morning briefing.
References (2)
- OpenAI, “Model guidance — Using GPT-6 Astra / Prompting best practices”: https://developers.openai.com/api/docs/guides/latest-model
- OpenAI, “Unrolling the Codex agent loop”: https://openai.com/index/unrolling-the-codex-agent-loop/
