architecture
One boundary for every write
Ask a security reviewer what they think of an autonomous coding agent and they will ask you one question back: what, exactly, can it execute?
In Agent Zero the answer fits in a sentence: only packages/runner may run a command or mutate a
checkout, and everything else must ask it through typed contracts. Working directories are
validated, escape attempts are rejected, timeouts and output limits are enforced — in one place.
The alternative is what most agent frameworks ship: shell access sprinkled across tools, each call site individually responsible for being careful. That is not auditable in any meaningful sense. A single boundary is. Your reviewer reads one package and knows the whole blast radius.
Isolation when it counts
Repository-supplied commands — your test suite, your linter — are the untrusted part, and those are
what isolation moves into an ephemeral sandbox. Every hosted lease has a maximum lifetime, quota
checks run before provisioning, and the agent only ever receives the ordinary Runner contract.
The evidence bundle records which runner ran what, so "verified in isolation" is a checkable claim
rather than a vibe.