Harness2 by Google Research – Recursive Agent Harnessing

Posted by irthomasthomas 3 days ago

Counter80Comment19OpenOriginal

Comments

Comment by rancar2 2 days ago

The readme links to the paper (in the badges), which is more helpful in understanding the project: https://arxiv.org/pdf/2610.14224

Comment by mchusma 2 days ago

I agree with what this paper generally says...but doesn't it seem like something anyone developing ai systems already do? Learn from past runs to update your skills and system prompt etc? I know I have done this a lot over the last year+.

Comment by jmathai 2 days ago

I think many of us have already started automating it so that they update alongside code changes.

Comment by cykros 2 days ago

Maybe I need to give this a more thorough read through, but it sounds largely like what harnesses, or at least, Hermes-agent, are/is already doing. That is. Learning from feedback and creating reusable skills in order to avoid so much reinventing of the wheel, in line with how the user's demands direct things.

Hermes is awesome and perhaps more should be mimicking them as they find paths forward, but is there anything different in kind here going on than they're already doing?

Comment by 2 days ago

Comment by endangeredhuman 2 days ago

They had Gemini CLI as open-source, used the community for validation, then moved to closed-source Antigravity CLI.

If the harness is closed, why open the "harnessing"?

Comment by ridgeComet617 2 days ago

Honestly the readme image bugs me less than the lack of a single end to end example you can actually run.

Comment by amritananda 2 days ago

Yeah, doesn't help that the "video overview" is just an animated slideshow of benchmarks. I guess that tells us who the intended audience is.

Comment by mobilejdral 2 days ago

The promo committee?

Comment by djc404 1 day ago

Bling bling ppl

Comment by agenticmode 1 day ago

[flagged]

Comment by tjpnz 2 days ago

>Harness² is a plug-and-play framework for recursive agent harnessing in an open world, where agents encounter tasks unseen during harness design.

Huh?

Comment by imafish 2 days ago

It's a plug-and-play framework for recursive agent harnessing. In an open world, agents encounter tasks unseen during the harness design. Harness² is designed for this exact scenario.

Comment by eithed 2 days ago

Your harness tells agent to do A when it encounters B. When agent encounters C and it doesnt know what to do it does D and you correct it to do E. Harness keeps track of this and makes agent do E next time it encounters C.

Comment by RamblingCTO 2 days ago

I always wanted to exponentially grow slop, thank you. The less I understand about what's happening, the better. Totally what I as a user want. Thanks G.

Comment by xyzsparetimexyz 2 days ago

I love that even google is using slop images in readmes now

Comment by 0gs 2 days ago

pretty convincing argument to be made that big G invented human web slop AND AI slop, full slop

Comment by chubot 2 days ago

Well I wouldn't say "invented", but I'd agree with "incentivized at scale until it became a cancer on the web"

The weird thing is that YouTube seems to have incentivized a lot of good content. Sure there is "scalable slop" like Mr. Beast, but there is also good content on YouTube (Fall of Civilizations podcast, etc.).

Somehow on the web I am not seeing a lot of good content incentivized by ads. I see a lot of WikiHow-type stuff.

It seems like writing moved to Substack and so forth.

Comment by hirako2000 2 days ago

That YouTube got great content, is that it has a quasi monopoly on video streaming.

Also the reason the noise ratio is high with junk, and now shorts.

Comment by lompad 2 days ago

Yep. They are the Slopfather, basically.

Comment by tohms 3 days ago

[dead]

Comment by behnamoh 2 days ago

[flagged]

Comment by embedding-shape 2 days ago

Well, as luck have it, Google don't let individuals be individuals if they work for Google, so anything under github.com/google or github.com/google-research could be a weekend hack project from an individual with zero association to Google itself, or the most supported project at Google, everything always go under Google's name if you work for Google. Basically, free-time as the typical person understands it, does not exists if you work for Google.

The disclaimer in the bottom of these project makes the distinction a bit clear, once you got to the bottom of the README:

> Disclaimer This is not an officially supported Google product.