I will share some actual data after a few weeks of IRL improvements, hopefully positive! ;-)

If a change seems worth trying, other agents draft updates and queue them for review and testing. The system remembers what was tried and what worked, so improvements build over time instead of resetting every run.

Those two views come together in a synthesis step. The system forms simple hypotheses like why a page might be underperforming compared to others. These are suggestions, not decisions.

Another part of the system looks at competitor pages covering the same topics. It records how those pages are structured and what they seem to focus on. This gives context for what might be missing or weak on our own site.

I built a small system that watches how a website performs. It reads traffic and behavior signals from PostHog and Google, but it doesn't change anything. It just observes and takes notes.

Oh! Haha. That *is* interesting to think about though!

Good luck with your project!

Just make sure you are storing those tokens securely if you go that route. Gold standard would be to use the system keychain if you can.

There are easier ways of giving your agent access to your inbox though. I am working on an app that requires inbox access and just found it easier to use app tokens for the major platforms that most people use. The apis are a useful abstraction on the hassle of managing email.

Oh, alpine is a command line email client. It started as pine, a variant of elm, and was the goto mail client on the early Internet before the web and TCP/IP came to Windows which drove gui clients like eudora and later gmail alpineapp.email