wtf is a loop?

A way to spend all your tokens faster and give your money to OpenAI, Anthropic and friends.

this is true

How is this any different than just saying " double check your work and test it before finalizing"? I understand that loops are for specific cases, are "designed" for verification, but at the end of the day it's just the AI agent verifying its own work one way or any isn't it?

you need a non-deterministic "loop" around the model to ensure it continues. it's not about double checking work, it's about checking against an end goal

Yeah... Like you said that's the hard part, most of the time I don't even know how it's supposed to work. I've been working on my characters to make a line queue to use a water cooler.. sounds easy "just make sure they all form a line".. It's been 2 days😅

In Codex or CC, is it feasible to have two loops running in parallel? One that analyzes output and gives feedback, raising the bar each iteration. The other that works towards the goals set by the first. Or better to design both inside one loop?

good question. I suspect one loop per goal is the right way.

This is a great summary. It’s surprisingly simple in concept (a trigger, and a goal) but the mechanism options (or even knowing how to define your goal) is where folk can get unstuck. Cheers.

yeah it was a lot simpler than I thought. still not easy to design loops though.

And the potential cost! The biggest elephant in the room…

the cost goes exponential with loops

/goal is the best loop

/loop is the best loop

I think this was a good video overall but I don't agree with the final speculation. Even in your scenario someone has to kick off the goal-oriented self-improving agent. There may not be someone sitting at a workstation directing it but that doesn't mean it isn't being pointed in a direction.

someone might kick off the initial loop, but if no humans are needed after that, it's self-improving.

@ThePrimeagen predicted this 😂

Recursive self-improvement is when a model can design a better model. A model designing a workflow that uses models is nested, not recursive.

Nice mat 🫡

Thanks for the explanation

thanks! a loop with immunity-agent is a dangerous combo:

Your point about feature development having amorphous goals (at 3:30) resonated with me. I haven't gotten pure spec-driven development to work well for complex features, mainly because I don't know exactly how the feature should work until I prototype it and then iterate on it.

i still didn't understand what is it loop

What is the tool you are using in the video?

Use Codex /goal with the product design plugin is another way to create a loop for a specific goal. Try that. I paused a /goal after 10 hours building a macOS app. Amazing results.

How this diff from Ralph?

Refine handles my loops.

/goal

Great video Matt

Matthew, you are the one asking the right questions! That’s why I follow you.

Bullshit on tokens becoming cheaper.

It is a fixed point iteration, not a "loop". Some are P-COMPLETE so unless P == NC there is no faster way to do it.

Just another reason why local AI will win.

But if we used a local model that will have more frequent running loops to increase output quakity through iterations, the limitation won’t be tokens… it would be time to market

thanks for always keeping us in the, ummm.., loop! :D

Copilot Cli has done this for a while now. Just: /autopilot goal

What page is that on Claude?

What’s the point of the loop if the prompt contains the goal? Just produce the goal.

Game changing tech. Great explainer!

The way I understand it: a prompt is one instruction. a loop is a small operating system around the agent. It decides what context to load, what task to run, how to check the result, when to retry, when to stop, and what to report back. The hard part is not making the agent do more. It is giving it a repeatable way to know what "done" means.

I swear some of the most basic shit in programming get you guys going. A basic programming concept and you're treating it like its rocket science.

Ok so we have a new term that will replace vibe coding now.

Thanks for this.