hey, first issue. i ran the same coding tasks through three claude models and never named a vendor, just to see what they'd reach for on their own.

payments is the one thing it never builds itself. it goes straight to stripe, every single time. auth it just writes itself, also every single time.

so it treats payments like a bank vault and auth like a weekend hackathon.

this stuff all works fine on your laptop, which is kind of the problem. your users aren't on your laptop.

my favorite one is caching. 36 runs and it didn't add a cache service once. it just writes its own every time.

the models really don't agree with each other either. sonnet never skips sentry, opus can't decide where to deploy, and haiku just wants docker.

so if you moved a project from sonnet to opus thinking it's the smarter one, you might've lost your error monitoring without noticing.

and opus likes postgres, while sonnet and haiku are fine with sqlite.

honestly i don't think the agent is wrong here. nobody told it the app needs to survive month three, so it builds for the demo.

next issue i'm testing whether having a provider's keys in your .env changes what it picks for auth.

sponsored

this slot's open. one labelled block per issue, and it never touches the numbers. whatagentspick.com/sponsor

nika