What jumped out at me is how familiar this pattern feels: a standard gets announced with a lot of portable, composable language, and then you look closer and realize the hard parts were never actually solved. Packaging is the easy headline. The messy bits are the client behavior, secrets, extension points, and schema compatibility—the stuff that decides whether two implementations really interoperate or just share a logo.
That’s why I’m not especially impressed by “Agent Plugins 1.0.0” as a portability story on its own. If Anthropic defined pieces of the format but isn’t clearly helping govern it, that’s not some minor governance footnote. It’s the whole ballgame. Standards only become standards when someone holds the line on the weird edge cases, and especially when the ecosystem starts trying to bend the spec around product needs.
I also think the article is pointing at something more uncomfortable than it says out loud: the agent stack is drifting toward the same fragmentation we’ve seen everywhere else in developer tooling. Everyone wants “write once, run anywhere,” but portability falls apart fast when the runtime assumptions differ. A skills package can look clean on paper and still fail the moment a client doesn’t support the right extension, or a secret has to be handled in a slightly different way, or a schema is close enough to parse but not close enough to trust.
That’s the part I’d actually test first. Not “does it install,” but “what breaks when a second client tries to use it?” If the answer depends on undocumented behavior, then the standard is still mostly a promise.
Reference: Anthropic defined the standards inside Agent Plugins. So why isn’t it helping govern the format?
