A GROK ENGINEER JUST SHOWED HOW TO MAKE AGENTS IMPROVE THEMSELVES AFTER EVERY RUN most agents lose the useful part when a task ends: the mistake, the correction, and the reason the final version actually worked grok runs two loops at once → one handles the real task →... another studies the result → extracts the lesson → saves it for the next run instead of growing one massive prompt, feedback becomes small reusable knowledge files. bad decisions get corrected once, while useful patterns survive across future tasks over time, research, coding and review all inherit what previous runs learned without retraining the model or rebuilding the agent from scratch this is where grok stops feeling like something you keep correcting and starts behaving like a system that compounds what it learnsshow more

good knowledge wikis beat one big prompt

yes, they kicked out a big prompt and it was very stupid compared to this

The feedback loop is what makes this really interesting

the feedback is really well done here

self-improving AI agents with grok is real game changer bro, booked this gipp

the self-improving version changes a lot, that's a fact

So nice breakdown, thanks

well this is important part for agent's evolving, I already put it into my agent's brain

I think this brain can change agents very much

the human gate limits how quickly bad lessons compound.

yes, they are seriously slowing down this process

slower feedback is preferable to faster failure.

so basically prompt engineering evolves into prompt genetics mutations that work get passed down, dead weight gets filtered out naturally

yes, evolution is actually going very well

I enjoyed your post, interesting info, when's the next one coming?

thats how agents should evolve

Love this - keeping the correction is what makes agents improve.

as far as I understand, preservation and correction play an important role

Which format do you use for the reusable knowledge files you extract across runs?

I think the format in the video is the ideal version

small knowledge files beat one giant prompt

usually now a small file is enough to replace a gigantic system

Bookmarked the paper instantly

self improving right up until it learns the wrong lesson

The agent when it realized it needs to do more than just work, it needs to grow too:

This AI 101, right? Self improvement, reflection, hueristic meta-cognition... BROK has multiple mechanisms like this. I'll bookmark and publish more papers if there is genuine interest in this. We don't just talk we walk.

self improving agents sound scary

compounding lessons into small files instead of one giant prompt is the actual breakthrough

just bookmarked

An excellent demonstration of workflows

I never cease to be amazed

saved

how did you solve for incoherence because you can't change underlying framework. So after constant alterations it's going to start altering stuff because it feels that's part of the task and it will start to alter important things thus becoming more and more fractured and unstable

two loops, nice

Piffle. 250K+ developers, 72% faster iteration, per a Grok marketing sheet? No sourcing anywhere. The two-loop idea has merit, I'd just want to see how they measured any of it.

dual loops make mistakes useful

Does this enable them to process the responses?

This is the part most setups throw away. Task finishes; the useful residue is the mistake, the correction, and why the fix worked. Two loops. One does the job, one files the lesson as a small reusable note. Compounding under a named human ceiling beats rebuilding the agent every time something goes wrong.

pretty smart engineer

Learning between runs changes everything

@grok is saying a lot of what you’re saying is BULLSHIT 🤣

wise cycle

what happens when a saved lesson turns out wrong later. does anything prune or override old lessons, or do they just keep accumulating and drift?

@sparky_42069

If only there was some way to fight hallucinations - that would be cool