Gobena Dache (ጎበና ዳጬ)

[Gobena Dache (ጎበና ዳጬ) @dache_gobena]

@cryptofergani Me

[Tezzo @Tezzo100x]

GM

[Gobena Dache (ጎበና ዳጬ) @dache_gobena]

@codewithimanshu Scalper

[Paulos Yibelo @PaulosYibelo]

Does harness matter? It CAN matter, A LOT. For this you don't have to look any further than ARC-AGI 1 & 2 solutions. for example, in the attached example you can see how Poetiq used a custom harness to improve GPT-5.2x-high by almost 20%, which is a massive delta for ARC.

What many don't realize is that the underlying models people already use either already come inside strong task-specific inference scaffolds and aren't just "strong models", and/or were trained from data generated through harnessed pipelines: synthetic task generation, candidate solution generation, execution, filtering, verification, ranking, refinement, and then training on the surviving traces. This means the final model is not separable from the scaffolding that produced its data.

This doesn't mean a model, out of the box, on its first continuation of tokens, is as good as what it would've been if something like Claude Code didn't have a strong harness around it. A harness changes the effective computation: it controls decomposition, memory, tool access, search width, verifier feedback, retries, self-debugging, and how failures, etc get converted into new attempts. Strong models are often a result of good harnesses, and good harnesses are often the thing that produces the data that trains the next strong model. Harness discussions can mistakenly imply that harnesses don't matter, or that harnesses only make models worse by constraining them. Bad harnesses absolutely do. But the opposite is also true, a pipeline with high-quality search procedures, technical priors, verifier signals, introspection, tool access, and domainnspecific decomposition can materially change output quality. Recent Mozilla and Chrome stable updates are good examoles.

Not all harnesses are good. In fact most I've seen rly suck, especially on the opensource side of hacking. It's super sloppy. Just like there are tiers of hacking skill, there are tiers of workflows, search procedures, and taste in what to try next. The transferability of those process tiers into AI systems is not surprising. Harnesses IMO need to evolve as model capability goes up so they don't kneecap the model. They should be built around weaknesses: bad long-horizon planning, weak verification, poor state tracking, shallow search, lack of environment feedback, and top tier expertise etc. Strong harness + strong model = even better results.

Will this always be true? Likely, but it's hard to say. Especially if you're a researchy type of hacker or novelty researcher, harness can mean a lot of things and can dramatically improve searchspace outcomes. A good model is a learned prior over internet-scale traces, but the prior still has to be queried, structured, expanded, tested, and selected from. Not only how you query, but how you structure intermediate states, outputs, failures, retries, and verifier feedback will significantly influence the result, just like how one human's subconscious process leads to discovering HTTP Desync as a class while another's doesn't.

I think if you don't have a really good harness, and aren't extremely familiar with the technical skillsets or top tier, have unique thought processes, and reasoning patterns that positively affect output quality, then it often makes more sense to have zero harness than to add one for the sake of it, which actually will hinder it.

[Gobena Dache (ጎበና ዳጬ) @dache_gobena]

@BitcoinIntelX BTC

[Gobena Dache (ጎበና ዳጬ) @dache_gobena]

@USronaldcarter Done

[Gobena Dache (ጎበና ዳጬ) @dache_gobena]

#NewProfilePic

[Gobena Dache (ጎበና ዳጬ) @dache_gobena]

#NewProfilePic