Why your agents fail (and why evals won't tell you)

Andrew Qu, Chief of Software in the Office of the CTO at Vercel, and Sanity co-founder Even Westvang join David Annez to talk about what actually works when building agents: reading traces instead of chasing a number, treating language as the material the machine is made of, and why Vercel deleted 30 to 40 custom tools the day a better model arrived.