Nathaniel Whittemore

OpenClock creator and now OpenAI employee Peter Steinberger wrote, Here's your monthly reminder that you shouldn't be prompting coding agents anymore.

You should be designing loops that prompt your agents.

On CNBC recently, Claude code creator Boris Cherny told the host that about six months ago, he shifted from writing code by hand to prompting Claude to write all of his code, but it turns out that even that isn't the limit of how things are changing.

Here's a snippet of a conversation with Boris from just last week.

Now this idea of loops isn't totally new.

Last year and into the beginning of this year, we had a lot of discussion around the Ralph Wiggum loop, which is basically a way to set up your coding agents so they would continuously try and try again without you having to sit there and prompt and interact every time they came up against the challenge.

Then in March we had Andrej Karpathy talking about auto-research, which was a specific type of loop designed to improve an actual AI model.

And now we've got the slash-goal primitive, which is a way that both OpenAI and Anthropic have embedded the idea of loops into the core experience of their major coding tools, Cloud Code and Codex.

The goal in each of these cases is to require less human intervention and get the AI to run for longer and longer, being able to fix its own mistakes, improve its own results, and ultimately accomplish much more comprehensive and complex tasks.

And so the point is, if you've already got people who are using agents opening up a significant advantage gap compared to people who are just using regular chat, and then you've got the vanguard of people at the labs who are even going a step farther in terms of how they're getting the most value, it strikes me as obvious that one of the places that the labs can try to democratize those experiences to more users is via the user experiences of their core app interfaces.

And that, I think, is the key point of the ChatGPT overhaul.

Now, yes, like I said, this does come with financial implications because the people who are running loops are burning way more tokens than the people who are just casually popping in to ask a question that they might have asked Google before.

I just think it's reductive to think that that is the primary motivation for the labs, as opposed to getting more people to experience the insane power and opportunity that comes with actually being able to run agents at scale.

And there is a lot of work to do.

RailwaysJustJake retweeted a post about loops and said, The problem with this and why I think people are frustrated, nobody has taught folks how to do this.

It feels both evidently the future and also somehow gatekept.

Just to be clear, he clarified, I believe it's being expressed at the fastest rate it can be.

It's just both evolving rapidly and so dense, like pulling a neutron star out of a magic hat.

Adding evidence to that, Seanu Matthew had a post go viral with over 300,000 views that retweeted Peter Steinberger's call to use loops and said, non-technical idiot guy here.

What does this mean for the non-coder audience?