[olegbk]

The boundary problem you're describing is also a trust issue, not just a connection lifecycle one. Even with ephemeral connections, you still need to know whether the agent making the call should be allowed to make it at all.

We use reputation-based admission control in production, agents below a certain trust threshold simply cannot invoke sensitive tools. The hallucinated tool call scenario you described is exactly what this prevents, independent of whether the connection is persistent or on demand.

agentveil.dev if curious.

[An0n_Jon]

Yeah the governance is also built into Orloj as well as applied policies. More so acting as guardrails at the runtime level. Ill take a look at agentveil and would love to know your thoughts on Orloj as well.