We gave the agent more access. That wasn't automatically better.
More permissions meant more confident wrong answers, delivered faster. The useful constraint turned out not to be what it could reach, but what it had to confirm before acting.

Short notes from building on Astronaut Industries. Things we tried, things that broke, things we decided AI should not be allowed to touch. Rougher than the essays on purpose.
Your business runs on three scarce things. They’re not interchangeable — and confusing them is the root of almost every AI failure.
Noticing. Someone asked, someone didn’t pay, something arrived, something stalled. Continuous — and no human can be continuous.
Knowing what happened before. What that customer ordered, why you dropped that supplier, what you promised in March. Most of it lives in one person’s head.
Deciding what something is worth, who to trust, when to break your own rule. AI doesn’t have this — and that gap is where businesses get burned.
Every note below is filed under the one it broke
More permissions meant more confident wrong answers, delivered faster. The useful constraint turned out not to be what it could reach, but what it had to confirm before acting.
Half of it collects structure the system can now infer. If nobody has to type it, the screen isn't a feature, it's a tax we kept charging out of habit.
Two lines of ordinary logic did the job with none of the variance. The discipline is noticing when the interesting solution is worse than the boring one.
An endpoint cheerfully acknowledged a field it silently dropped. Everything downstream looked fine. Anything that acts for you has to be loud when it doesn't.
A commerce engine with no volume pricing at all, and coupons with no minimum quantity, so a code meant for two carts fires on one. The strategy moved somewhere nobody was looking.
An add-to-cart returned false and the handler just returned. A dead control that looked completely healthy. Silence is the one thing an operator can't debug.