Separate proposal from deployment
An agent may identify a weakness or propose a new skill. That does not authorize it to change a production system. Keep generation, sandbox evaluation, and promotion distinct.
Benchmark the actual mission
A useful evaluation reflects domain constraints, failure conditions, and evidence quality—not only a general score. Review regressions as carefully as improvements.
Make reversibility concrete
Version the agent, its tool permissions, evaluation results, and approval record. Preserve a fallback and define when rollback is required.