Are you saying they did not use native harnesses like Claude Code or Codex? How did they do it then?

In the last three startups I worked at I didn’t bother exercising my vested equity - even a successful exit would at best triple the price of those shares - not worth the risk. One of those three startups already failed.

I’m forced to do this with claude code: documenting work for context management. Every new agent starts fresh, so everything better be recorded and explained.

I’m observing very similar trends at a startup I’m at. Unfortunately I’m not ready to retire yet.

I think it’s a combination of simulation, YouTube videos, and specially recorded training footage. The last one is expensive, but given the funding these startups receive, I’m pretty sure they can scale their RL methods at least 10x.