π a16z: Knowing When to Stop β The Art of Making an AI Loop Converge
a16z partner Yoko Li unpacks "loop engineering" β letting AI agents iterate on their own work until a verifier says stop β and why most loops don't know when to quit.
β’ In one live test, the first $1.40 of spend took a Lighthouse performance score from 26 to 89 β the next $2.84 (67% of the total bill) bought zero additional points.
β’ A web-agent benchmark found going from 1 to 10 samples lifted success from 38.8% to 43.2%, but doubling again to 20 samples added just 0.2 points for twice the tokens.
β’ "The systems that matter will not be the ones that can keep going. They all can," Li writes β the differentiator is deciding in advance what "done" costs.
π Read more
#a16z