An ibl.ai agent can now be watched doing real work from inside its sandbox: in this three-minute video it calls a live external API through a host the organization allowlisted, with a secret it can use but never read, and on ibl.ai you own all the code and the data, the sandbox included.
ibl.ai: AI agent calls a live API from a sandbox (2:58). Also on YouTube.
What does the video show?
The agent runs in the sandbox type that shipped in September: a real Linux virtual machine bound to the agent, with no network by default. The video walks through the three controls that let it reach the outside world safely, then shows the call happening.
- A network policy names the one host the agent may reach. Everything else stays closed.
- A secret holds the API key. The agent can use it in a request; it cannot print it, read it or send it anywhere else.
- The call itself runs from inside the sandbox, and the result comes back into the conversation like any other tool output.
Why does it matter?
An agent with a real machine is useful. An agent with a real machine and open internet access is a risk no regulated organization will accept. The sandbox gives you the first without the second: the organization decides, host by host, what its agents can touch, and every call is logged.
The controls in the video are the ones described in Agent Sandboxes: A Real Linux VM, Locked to Hosts You Allow and Agent Network Policies and Secrets, Now in Settings. They are part of Agentic OS, and like the rest of the platform they run inside your own perimeter: on ibl.ai you own all the code and the data.