Self improving product loop, the holy grail!
Core things you need to make this work:
- REALLY good "preview environments" -- invest in the last mile in being able to run your system full stack in the cloud, per git branch (maybe i'll write about how we did this at some point)
- Observability in prod (easier than the above)
Agents are good enough to read prod traces and seed these preview environments with data that reproduces agent behavior! Once you do this, your coding agent can iterate against that "eval".
Majority dynamics on sparse random graphs, with Ashwin Sah: paper, proof atlas, full Lean formalization.
N people hold random opinions; one side leads by ~√N. Each day, everyone adopts the majority among their neighbors. With average degree N^ε, that lead wins in O(1/ε) days.