How are you running AI agents on air-gapped or closed networks?

Banks, hospitals, defence and government teams want AI agents too, but their networks can't call out to a cloud sandbox, and the agents can't hold real credentials.

We've been building kosa8 for exactly this: every agent gets its own microVM with its own kernel, the network is closed by default, keys stay on the host, and it runs fully air-gapped. We published a technical report on how we tested it: kosa8.com/research

Before we launch on Product Hunt on Nov 3, we'd love to hear from people working in closed environments:

1. What kinds of agents are you running or blocked from running (support, data, document, SOC, coding)?

2. What does your security team need to see before an agent can run unattended?

3. How do you handle model access and updates when nothing can leave the networ?

@fmerian @jesv curious what you've seen from teams in regulated spaces.