Hi everyone, we're running into a governance issue with a multi-agent setup and wanted to check if others have faced the same challenge
Our setup : there's one orchestrator agent (published, accessible to all users) that delegates tasks to several specialized sub-agents via the run_agent tool. These sub-agents are not meant to be used in isolation, they only make sense when called by the orchestrator
The problem : we want sub-agents to be invisible to regular users (so they don't stumble upon them), but if we keep them unpublished, the orchestrator can't call them on behalf of regular users. If we publish them, they become visible to everyone (which creates noise and potential misuse) -> which is therefore the current set up of the orchestration, because we have no other solution for now
We considered restricted spaces, restricting tools/KNB to a skill and not the agent, but that does not really solves it (we can't rely on skills exclusively)
Two questions :
- 1.
Is this a confirmed limitation of Dust's current permission model?
- 2.
Have you found any workarounds to handle this?
Thanks in advance for any feedback or REX!