I’ve been thinking about a problem around sharing specialized AI systems.
Suppose a domain expert builds an agent based on their own knowledge, workflows, prompts, tools, and evaluation methods. They may want other people to use the capability, but they probably don’t want to expose everything that makes the system work.
So I’m wondering if there’s a useful separation between access to a capability and access to its implementation.
Ideally, the creator could control things like who can use it, permissions, versions, updates, and revocation, while users interact with the capability through their own accounts.
I don’t think this can provide perfect protection. If users can interact with a system, there will probably always be some possibility of learning about its behavior. So I’m more interested in a defense-in-depth approach: isolation, permissions, monitoring, evaluations, versioning, rollback, etc.
This makes me wonder whether something like this could eventually become a broader AI capability distribution layer, rather than just sharing prompts or custom agents.
I’m still learning the technical side of this, so I’d really like feedback from people who have actually built agent platforms or dealt with protecting proprietary AI workflows.
Is this a useful architectural direction, or am I missing an important limitation?