Satya Nadella posted a few days ago on X about the potential risks of Artificial Intelligence (sorry — “Super Intelligence”).
The crux of his piece is that while we can now draw a cheap supply of intelligence from a tap, we still need frameworks to manage its authority. Zero trust principles should be applied to models, and the only way to trust agentic workloads is to place as little trust as possible in the models themselves.
Frontier models provide intelligence on demand, but you still need infrastructure around them to:
- Provide a scoped list of tools they can access.
- Guard against hallucinations and ensure they use the right tool for the right job.
- Provide guardrails so users can only ask models to do work they’re authorised to request.
- Provide transparency into how models approach tasks and execute work.
It’s a timely piece because the iviva Agent Harness ticks all these boxes and is specifically built for the CRE world.
- Guardrails — The harness provides guardrails through iviva’s well-established and extensive role-based access controls.
- Diversity — Given iviva’s enterprise deployment options, we often need to deploy in highly secure, air-gapped environments. This means the harness needs to work with more than just frontier models. iviva’s harness has been tested with a variety of open-source models that can be deployed on-premises, including Gemma and Qwen.
- Independent Controls — Since all actions a model takes are routed through iviva’s secure layer, the harness can automatically request confirmations and approvals from users without relying on models to decide when they’re needed.
- Containment — Agentic workflows can be terminated mid-task and shut down completely if required. Without the harness coordinating activity, the model has nothing to rely on for its work.
- Auditability and Disclosure — All agentic workloads are auditable: which models were used for each task, what code was written for each request, how tools were used, and what confirmations users gave.
One of the side effects of creating zero-trust agents is that the models actually perform better. The less “busy work” you ask them to do, and the more that gets lifted into the harness, the smarter they are and the faster they perform.
If you’re looking to bring intelligence into your properties, talk to us. We’ve just launched our AI Accelerator program, and you should get in touch!
