Nvidia demo shows AI coding agents build faster with DOCA 'skills'

A demonstration from Nvidia compared two AI coding agents tasked with building a Go-based RDMA application on Nvidia's BlueField-3 platform using DOCA software. One agent had access to Nvidia's curated DOCA 'skills,' while the other had to figure out the needed modules and build steps through trial and error.

Both agents eventually produced working applications verified on real hardware, but the agent using DOCA skills got there with about 73% less handwritten code and 50% fewer hardware interaction commands. It also converged faster on Nvidia-supported sample code rather than reimplementing low-level functions.

The demo is meant to show how giving AI coding agents structured, domain-specific knowledge can speed up development of complex, hardware-level networking software while keeping results accurate and verifiable.