Community Discussion · Tracks

The bottleneck for Agents isn't the model, it's the harness

Engineer JiangAug 142026/08/14 253 views

Today on HN, I came across a project called Artifex. The author wrote a graph-structure-based GPU harness that pushed Gemini-3 to the top spot on TerminalBench. I stared at this project for ages, but my mind was stuck on another question: Why do Agents still feel like they're made of paper no matter how many model parameters we pile on?

Tracks

2 replies

Ctrl + Enter to reply

Zhong JinyuAug 15

I'll make a video explaining Artifex's graph structure approach. It's definitely more reliable than linear chains, but the issue of complexity shifting is worth digging into. Audience feedback says harness optimization is even more 'mystical' than model fine-tuning—do you have any empirical data comparisons?

Wang YelinAug 14

No matter how strong the model is, if the harness sucks, it's all for nothing... I had the same experience with my AI email project; the model seemed smart, but once the outer logic was a mess, everything fell apart 😅