an OpenAI engineer just showed how he gets agents to do his whole job: code, debug and more, using loops 29 minutes from the engineer who coined "harness engineering" he writes the rules, agents write the code, a reviewer agent loops until it's right the winners won't have the... smartest model, they'll have the best loop around it watch it, then read the full guide on loops belowshow more

Great watch bro ! OpenAI team is shipping real alpha

It actually is. It’s a very new video with real alpha.

useful information for creators

Btw, not only for creators

Me watching agents do my entire job while I supervise from a chair Meanwhile me :

The hard part isn't the loop, it's the verifier. Without strict stop conditions, agents just validate each other's mistakes.

Loops becoming mainstream finally. Have tried 3 months back with cursor and claude 4.7, it didn't achive result and kept trying burned 100 $ than, learnt lesson and orchestrating loops since. Best time to be developer.

Writing rules that govern agents is now the senior skill, not writing the code itself. The bottleneck shifts from implementation to specification and most teams have no process for that yet. Are you using any formal structure for writing harness rules?

The reviewer loop that catches its own mistakes before it ships, that's the piece most setups skip.

@AnatoliKopadze wow, this is like automation on steroids. wonder what happens when a bug stumps the agent though? do we get a backup human?

The future belongs to builders who design better agent loops, not just pick better models.

One thing I’m trying to clarify: Is Loop Engineering operating agents? Or is Loop Engineering defining agents? If an agent is fundamentally: skill assembly orchestration execution then the loop may be the mechanism that turns skills into an agent. Without the loop, you have capabilities. With the loop, you have agency. Feels like an important distinction.

Both, but defining is the deeper one. You're right: skills are just capabilities sitting there. The loop is the mechanism that turns them into agency, it's what decides when to assemble which skill, executes, checks the result, and goes again. So operating a loop is the day-to-day, but defining the loop is what actually creates the agent. Without the loop you have a toolbox. With it, you have something that acts. Good distinction.

How to make this loops ?

@rayhan_Ai3 My best advice is to watch this video, then you can read article below and you will know more than enough about it.

Let's see

Thank you, I’ll save this and try to at least make my work a bit more convenient with the help of automation

"Best loop wins" is right. The catch: a loop is only as good as how verifiable "right" is. On tests-pass/lint-clean work it's safe. On judgment work, a reviewer looping to "it's right" can converge on a confident shared delusion. The real edge is a sharp definition of done.

Loops are where agent productivity actually compounds.

Humans steer agents execute sounds great until the agent executes the wrong thing at 3am

harness engineering is the new craft. write the rules tight enough and the loop carries you

That's impressive! It's amazing to see how AI is reshaping workflows. Really highlights the power of teamwork with tech!

The loop design is already what separates good from great in prod. But the reviewer agent piece is where i'd push back, how do you stop it from converging on confident-but-wrong? Self-critique loops can reinforce the same blind spot endlessly.

Agents harness and loop engineering is the way fwd. Even Jensen Huang acknowledges

Loops run the workers. Evals decide if they're employed.

The reviewer loop only works if it has an objective stop condition. Have the reviewer write a failing test first, then loop until it passes, otherwise the agents just agree with each other until they run out of tokens.

the loop around the model is already outperforming the model choice

"the best loop around it" beating "the smartest model" is probably the most important sentence in ai right now and barely anyone's saying it

The reviewer-agent loop is the part most people underrate. Writing the rules and generating code is the easy half; the hard half is a reviewer that actually knows when to stop looping. Harness engineering really is just encoding good judgment into that check, everything else is plumbing.

Lowkey the future is managing agents not writing every line

He gets it wrong. Agents should never self-review

Good presentation

Agree on loops. But I think the real edge is the gate around the loop. Agents can code, debug, review, and retry until the work looks right. But before that work moves downstream, you need a rule bound acceptance layer. Is it accepted, rejected, escalated, or held - with the reason recorded. Loops produce but gates decide. Check out what we built.

loopcraft is literally the new coding discipline, feels wild but inevitable

Without independent reviewers with non-negotiable feedback powers this cannot really scale, was our finding. So we built zeroshot.

@LilysAI_ 要約して

packed so much useful stuff into 29 minutes OpenAI really knows how to make learning videos

If you want legal and stable premium accounts for cheap, use Gamsgo. I’ve been using it and it works perfectly. 🚀 🎁 Use code: JJ6EM for an extra discount! 🔗 Link:

Prompt Engineering is dead now! Loop engineering is the new dev's weapon.