Founder and lead engineer, Mynd Labs
Markdown supported · ⌘↵ to post
This is the right framing: the failure is that the output looks fine while the permission model is quietly wrong, which is the hardest kind of bug to catch because nothing actually errors. Modeling a request as an action with a purpose, allowed sources, and recipient constraints is what I have found actually scopes an agent, far more than handing it the whole workspace and hoping. In the four-layer model, which layer ends up owning the deny decision when purpose and identity disagree?
Most agent bugs I hit are really tool description bugs. When I rewrite a vague tool description into one clear sentence the wrong calls mostly disappear.
Yethikrishna R
Vlad Zoff
I'd also carry those constraints through to the output, not stop at retrieval. An agent may be allowed to read two sources separately but not combine them into a summary for a recipient who can only access one of them. So the policy needs to cover both source access and destination constraints, including derived content. Otherwise narrow retrieval can still result in a broader disclosure.