GPT‑5.1-Codex-Max’s ‘24-Hour Agent’ Promise vs. Reality issue
OpenAI says that GPT‑5.1-Codex-Max can run nonstop, automatically compacting its context so the agent keeps working until the task is done. They even mention it can work independently for over 24 hours.
But in my testing—since launch until now—that hasn’t matched reality at all.
Every time I give the agent a feature to implement, it never completes the task on its own. Instead, after just 1–5 minutes (and using only 1–5% of the context window), it stops and suggests “next steps.” I then have to manually say “continue,” and it runs another short burst—again stopping after a few minutes. This repeats indefinitely.
After about 3 “continue” prompts, it often drifts off-track: the “next steps” it suggests are outside the original scope, unrelated to the feature I asked for, and we lose focus on what we were actually trying to build—even though I explicitly told it: “Do not stop until the task is fully complete.”
So I’m wondering:
Is the “24-hour autonomous agent” claim just marketing hype… or worse, misleading?
For reference: https://openai.com/index/gpt-5-1-codex-max/
Model used : GPT‑5.1-Codex-Max Extra High
7 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
The behavior you're describing doesn't generally match what we're seeing. We have had some reports of "stubbornness", but it sounds like you're hitting this early and often. Let's see if there's something specific about your configuration.
Can you provide more information about your configuration? Are you using a custom AGENTS.md prompt? Do you have MCP servers configured?
Do you see the same behavior with models other than "max"? Or with reasoning effort other than xhigh?
There are no MCP servers used, I manually disabled all of them in the config file. I was working on refactoring code from one language to another. The codebase is large.
I’ve tried running it both with and without a custom
AGENTS.mdfile, and the behavior is similar in both cases.The expected behavior is that when the context window limit is reached, codex should automatically compact the context and continue processing. However, it currently uses only up to 20% of the available context window and then stops and suggests next steps
I only tested "max" with xhigh for refactoring
Thanks for the additional details. This sounds like model behavior, not something specific to your configuration. You may be able to mitigate it by adding additional instructions in your AGENTS.md or in the prompt itself (e.g. by telling it to "be persistent" and "don't stop until the problem is solved").
If you have a session that is a good example of the behavior you're seeing, could you use the
/feedbackcommand to upload it and then post the thread ID here?Ive been using codex / claude code for a while.
I am getting the same bug here today.
I instructed it not to stop until the task was complete, but it ignored that instruction.
OK, we'll share the feedback with the team responsible for training our codex models. If you see other model behavior that you'd like to report, please use the
/feedbackcommand. That will provide us with additional details including logs and conversation text.