Take a long road trip with Auto Drive 🏎🏎🏎

[zemaj]

Most experiments I run with Code, GPT-5, and Codex never leave my laptop. They’re interesting, but they don’t actually make the experience better.

Auto Drive is... different

Here’s the problem: GPT-5-Codex is so good that half the time I’m just typing “please continue.” It’s powerful, but it wasn’t really built to manage long-running, multi-step goals on its own. That’s where Auto Drive comes in.

Auto Drive lets Code tackle tasks on much longer time horizons, with built-in fault tolerance. It’s basically GPT-5… driving GPT-5-Codex with some guardrails. You hand it a high-level objective, and it just keeps going - planning, retrying, pushing forward - until it’s done (or until you stop it).

And the results so far have blown me away.

Think commands like:

/auto make $100  
/auto find the most urgent upgrade and code it  
/auto clone reddit.com  
/auto update the screenshots on my website  

Then just let it run for a day or two.

It’s still experimental. It might crash or go off the rails. Multi-agent support and other upgrades are on the way. But the early version was so impressive that I couldn’t sit on it - I had to get it into your hands immediately.

Give it a try. Push it hard. And tell me what you think.

[enieuwy]

Wow. Incredible work!
Very excited about this (and this project in general)!

[zemaj]

Have introduced a few new improvements to Auto Drive. It does still sometimes end up in loops when using broad instructions. Have some upgrades coming soon to stop that.

I like testing the /auto make $100 prompt. Not because it's a good prompt (more specific guidance is best), but because it is a bad prompt. The system should be able to interpret ANY prompt and work consistently within the constraints it has.

If anyone DOES actually get it to make money autonomously, please let me know!

[zemaj]

Just ran it for over a day and completed this massive refactor in one shot #303 😳

[ytwytw]

It would be nice if I could plan first and then complete it with auto mode.

also the conversation resume/continue still have issues with coder (cannot continue, context error, etc)
if the issue is fixed, and resume and continue with auto mode, that would be goat

[ropl-btc]

yeah good point. i guess you could use plan mode first and then go ahead and execute it with /auto - not sure if that works well? cc: @zemaj

akin to how deep research also asks you for some confirmations before it goes to do its thing.

[kalbur]

This is insanely cool. One thing I've noticed is on the codex cloud agent, it doesnt change over to the working branch from the terminal. It just defaults to the main branch?

[enerage]

Best feature! Props to you for developing this. This solves the biggest pain for codex, I just wish it was a bit more stable now, sometimes it just hangs for me, and I haven't found an easy way to resume/restart from where it hanged.

[zemaj]

Yeah it's a bit buggy right now. But there's a massive, massive upgrade coming in the next couple of days. It significantly improves auto drive's performance, reliability, agent use and so much more. Turns it from an interesting experiment to a real autonomous system that can do many real world tasks. Keep an eye out for it!