← All work
Agent Runtime & Isolation

Running untrusted agents in production

gVisorCilium eBPFKubernetes
Runs in production behind a customer-facing AI product.

The problem

The moment an agent can run code, write files, or call an internal API, it stops being a feature and starts being a workload with your permissions. The industry's default answer has been to restrict agents until they're safe and useless. The honest answer is that a standard container is not a security boundary — it shares a kernel, and a kernel exploit is the whole game. If you want agents doing real work against real systems, isolation is the product, not a checkbox.

What I built

An agent runtime platform and Kubernetes-based sandbox with two layers of enforcement: gVisor intercepting syscalls in userspace so a compromised workload never talks directly to the host kernel, and Cilium eBPF enforcing network policy at the sandbox boundary so egress is allowlisted rather than assumed. Built for production coding agents and AI-native workloads, not for a demo environment.

What was hard

Isolation costs latency and developer patience. Every boundary you add is a boundary a team has to understand before they ship. The design work wasn't picking gVisor — it was making the secure path the default path, so a team onboarding didn't have to become kernel-security experts to get their agent running.

What this means for vendors

Teams shipping agent features do not want to run this themselves. The category is converging fast — managed agent runtimes from the hyperscalers, ephemeral sandbox products from the edge platforms — and they're all competing on the same axis: how little does the customer have to know. Whoever makes strong isolation a one-line default instead of a research project owns this layer. Everyone else is selling a kit of parts.

Platform operations as agent-executable skills →Rules for the 80%, models for the tail →Right-sizing a fleet with a seasonal load curve →Shipping generative analytics before the playbook existed →Grounding a sales assistant in renewal knowledge →Building the operational view from zero →Scaling content governance with policy and data →