Add idea: stronger workload isolation for agent workloads

ideas/stronger-workload-isolation-for-agents.md

@@ -0,0 +1,29 @@
+---
+title: Stronger workload isolation for agent workloads
+author: Rashid Rashidov (@rrashidov)
+date: 2026-07-08
+tags: [sandboxing-isolation, runtime-lifecycle]
+---
+
+## The idea
+
+Agent workloads run instructions chosen at inference time — the code that executes is not
+something a developer reviewed. A shared kernel with other tenants is a meaningful attack
+surface that traditional app workloads don't expose in the same way.
+
+CF should offer a **gVisor-backed runtime option** for apps that need stronger isolation.
+From the developer's side this would look like a manifest field — something like
+`runtime: sandbox` — that opts the app into a user-space kernel boundary without any other
+change to how the app is pushed, scaled, or observed. Operator policy could make it the
+default for orgs that host agent workloads.
+
+The design of the enforcement point, and how gVisor (or an equivalent) is wired into the
+container lifecycle, is open for the platform experts to propose.
+
+## Related
+
+- `research/k8s-agent-sandbox.md` — documents how k8s-sigs uses `runtimeClassName` to
+  decouple the isolation mechanism (gVisor, Kata) from the workload API; the pattern is
+  directly relevant here.
+- Companion ideas that build on this foundation:
+  [[localhost-only-egress-for-agents]], [[credential-less-agent-processes]].