Run browser-agent experiments overnight. Review the results in the morning. StackAI CTO Frank Hidalgo, PhD used GPT-6 Astra in Codex to do just that. The team used Asana’s Command to turn findings into changes that shipped. The team tested four models across 144 runs. The optimized GPT-6.1 Sol workflow delivered 76X lower estimated model costs and roughly 5X faster runs than the original production setup. Read the full story: https://lnkd.in/eVqki_As
144 experiments later, someone finally yelled: DUDE, YOU WERE ON THAT PAGE FIVE MINUTES AGO. (using all caps for historical accuracy)
76X lower model costs from testing four models across 144 runs. That's a good case for treating model choice as an experiment, not a fixed decision. Running it overnight and reviewing in the morning is a smart way to work.
76X lower estimated model costs and 5X faster runs is impressive. A great example of combining AI automation with practical business results.
Finally I can let these browser agents grind overnight without waking up to a heart attack from my API bill. 😂