Agent Isolation Boundaries

An agent isolation boundary defines what a coding agent can see, change, and contact while it works. The research compares several boundary types: same-process approvals, operating-system sandboxes, containers, virtual machines, hosted machines, and embedded runtimes.

The strongest practical boundary in this research is a separate virtual-machine kernel, as seen in Clawk, Firecracker, and Apple Container. A container or policy wrapper can still be useful, but it shares more host assumptions. The main effect is simple: a stronger wall lowers host risk, but usually costs more setup time, memory, and integration work.

Sources

• • source analysis.
• • source analysis.
• • source analysis.
• • source analysis.