@system_monarch I would redesign around different trust, different blast radius. Like give humans, ci and agents separate identities with scoped, short lived credentials. Then enforce different permission and spend limits based on identity or workload.
After years of working with and reading about databases, started my own little LLVM-inspired database in Rust. Everything from scratch, with one rule: zero vibecoding. Should be fun.
This is how ir looks like on day1
Don’t get out of the loops.
Stay in the loop. Stay sharp.
We’re going to face a lot more AI-generated junk, and no amount of loops will magically fix bad engineering.
Most engineers optimize for average latency. Users remember the slowest request. A 20ms average means little if your P99 is 2s. At Amazon’s scale, those slow requests often hit your most valuable customers. Optimize the tail, not the average.