Two Walls Around an AI Agent

Everyone in Oslo arguing about AI sovereignty is arguing about where the servers are. That is the outer wall. It is the half of the question that gets asked.

The other half is what your agent is carrying. Secrets leak into context, into prompts, into logs, into a transcript that goes somewhere you did not choose, and they do it long before anyone gets near the machines those secrets are stored on. That is the inner wall, and it fails quietly.

On 22 September we put both walls in one room. Two of the engineers building the outer one, and one who spends his time on the inner one. Nobody on this bill thinks you get to pick just one.

What We'll Explore

• What an agent harness actually is, and why it turned out to be the thing everything else hangs off • Where secrets actually leak: context windows, prompts, logs, chat history, and someone else's servers • How to keep secrets encrypted in git and still let an agent use them. SOPS, key management, and handing a value to a process without it landing in a prompt • When owning the hardware is worth it, and when you can happily let the big models run on someone else's, if the inner wall holds • What each of them would not put inside an agent's reach, and why • How OpenGeni went from nothing before the summer to what it is now, told by the people who wrote it

Speaker(s)

*Nikolai Norman Andersen* is CTO at Variant Oslo, and in his own words gives practical technical talks about secure, privacy-aware software delivery, with a particular interest in infrastructure, developer experience and handling secrets. He brings a revised version of "*How to Keep Secrets from your Agent*" from JavaZone earlier this month: how to store secrets encrypted in git, and how to let an agent use a secret without ever seeing it. Expect SOPS, key management, and stdin-based handoff patterns that keep plaintext out of prompts, logs and chat history.

*Jørgen Kristiansen Sandhaug and Bendik Nyheim* are two of the people building OpenGeni here in Oslo, an open-source sovereign agent harness: autonomous agents on infrastructure you own, any model, durable sessions, human approvals, MIT licensed. Jørgen started the project before the summer, out of earlier work at cloudgeni.ai. Bendik came in as a founding engineer. They take the arc together, from the first version to what it is now.

Who Should Come

Builders running agents against real systems, architects deciding what an agent gets to touch, security people who have been asked "is this safe yet", and anyone who has been handed the sovereignty question without a way to answer it.

You do not need to have picked a side on where your models run. Not having picked one is roughly the point.

Bring your curiosity and your beverage of choice.