Two engineers, two answers to the same question

The interview I read on Ars Technica opens with exactly the kind of line that stops me: Anthropic's Claude Code product lead, Cat Wu, says her team deliberately chases a "lean harness" — they don't build an opinionated structure around the model, because the model itself changes within weeks. Augment Code's Vinay Perneti argues the opposite: in a large, private codebase, indexing context ahead of time beats making the model rediscover it with grep on every task.[1]

I'll confess I think they might both be right — because they're not actually measuring the same thing. Wu found that a handful of language server protocols didn't measurably improve performance; Perneti's harness does something much broader. But they converge on one point: frontier models will keep improving at this exponential pace for at least another year.[1]

As the model layer gets cheap, the harness is what wins

The most striking thing Perneti said was that frontier models like Opus or Fable could get too expensive to remain the only option in everyday agentic workflows — which would push smaller, open-weight models into wider use. Reading that reminded me of what happened at OpenAI the same day: Dean Ball argued for stoking "regulatory fear" around open-weight models, then backed down after pushback from Yann LeCun and Hugging Face's Clem Delangue. The two stories are really two faces of the same tension: as the model itself commoditizes, who gets to control that commoditization?[1], [2]

The model layer is commoditizing, and value is moving up — I wrote the same thing on July 19 about the agent layer getting cheaper, and today's news sits on the same line. We're now looking for the difference in the harness, and in who gets to access the open model.[1], [2], [3]

The next front: the harness market

On July 19 I wrote that an agent had been built for under a thousand dollars, and that the model layer was commoditizing; today's debate confirms it. My bet is that the coming months will bring a real market test between Anthropic's "lean" approach and Augment Code's "context-rich" one — and whichever wins, the developer's actual job will stop being about picking a model and become about building the right harness around it.[1], [3]

If Perneti is right and open-weight models gain ground as frontier models get pricier, the regulatory-fear idea Dean Ball just retracted could resurface — this time with a stronger case behind it. The chef keeps handing out tasks to the agents; but which kitchen she's using now matters almost as much as the recipe itself.[1], [2]