NVIDIA sells the compute that enormous models run on. So it caught my eye when NVIDIA’s own researchers published a position paper arguing that small, specialized language models are the better fit for most of what agents actually do. You have to admire the honesty.
The researchers don’t say general-purpose models are unnecessary, and the paper deserves that precision. They propose mixed systems: small specialists for the repetitive work, a large model kept where open-ended conversation matters. Fair enough. But even that careful position walks away from the dream of one model that does everything.
The build-side evidence points the same way. One benchmark handed a model every available tool description, and it picked the right tool 13.62 percent of the time. Retrieve only the relevant tools and accuracy jumped to 43.13 percent, with prompt tokens cut by more than half. Tripling accuracy by removing tools… that’s the whole point.
Every engineer actually building agents learns this fast. Limit what goes into the context window. Limit the tools. Clear the context on a schedule. A clean, targeted set of tools gets you what you want far more often than a super-powered model with everything bolted on. You don’t want a dog that walks on five legs.
So before betting the roadmap on the next bigger model, it’s worth trying subtraction first: fewer tools, a smaller context, a scheduled wipe. The next real improvement in your agent program is a subtraction, not an upgrade. What have you removed from your agents lately?
Originally posted on LinkedIn.
RationalAI helps organizations turn ideas like this into results: AI strategy and advisory services. Previous article: Approve the Model. Trust the Harness.
