Trust, but Sandbox

Trust, but Sandbox

Agents are becoming more autonomous, are expected to work more independently, and are being granted greater privileges. They install tools as and when they need them. And they can access all files on the system, use the browser with existing cookies, and much more. How can we allow the agent to work while still maintaining control?

The talk will present three methods for isolating coding agents and compare their advantages and disadvantages. First, I’ll outline the capabilities of a local agent and the risks that come with them. In a live demo, I’ll show just how simple sandboxing can be and how great it feels to switch to ‘YOLO mode’ without a second thought and let the agent work on his tasks. Finally, we’ll take a look at some practical features that a separate environment for agents can offer.

Michael Krämer is working in software development for over 20 year and works as a Software Architect at INNOQ. He is very commited to work out component designs with clearly defined responsibilities and to find appropriate solutions for requirements. Besides that he is engaged in Machine Learning, the integration of ML models in production environments and works as a trainer for Software Architecture.