A cadence of long-form writing on the principle and its applications.
The architectural reason why most "self-improving" AI agents in production today are not, in fact, self-improving — and the explicit, auditable boundary that turns the pattern into something an SLA can be written against.
Five layers: kernel control loop, self-observation and trace, mutation operators, gating and rollback with the jurisdictional firewall, and continual evaluation. Each one's responsibilities, each one's invariants, each one's failure modes.
The line between recursive agents that compound value and recursive agents that compound entropy runs through the eval suite. What capability evals catch, what operator evals catch, where each fails, and why the agent must never have write access to either.
Notified when the next essay publishes.
One email per essay. Roughly every 14 days. No other sends.