What we have actually learned building and running these systems, written out at length. No gated downloads and no email wall. If it is useful, take it.
Everyone is tuning prompts. The gains came from the environment built around the agent since Claude Code launched in February 2025: the standards it reads, the gates it cannot fake, the memory that means it does not relearn the codebase every morning.
People over profit, treated as an engineering decision. Structure a business so the profitable path and the ethical path are the same path, and the incentive structure holds the line instead of willpower.
The papers here are the long version. The shorter, more frequent notes (what broke, what it cost, what he would do differently) go out on LinkedIn first.
