FAQ¶

How does it differ from an agent-native sandbox? An agent-native sandbox answers whether it can block something. On top of that, pVisor provides retrospective selective merging, conflict protection, and a checkable execution record, with consistent semantics across agents and executors. See comparisons.

Does it slow agents down? pVisor adds the cost of admission, file interception, and recording. Measured startup and filesystem overhead is in benchmarks and comparisons (partly under construction).

Can it run in CI? Yes. You can use it in a pipeline the L1 way today: let the agent finish, review, and merge only what you want. Policy- and evidence-driven exemption and clustering (L2/L3) are the direction; see running agents in CI.

Which agents are supported? Any command can run; Claude Code, Codex, and others get presets matched by executable name. See connecting your agent.

Does data leave my machine? pVisor itself does not collect or upload usage data. But an agent's calls to model APIs leave the machine anyway, and pVisor does not keep them local: under --safe, only that agent's model API allowlist is permitted and other egress follows policy. pVisor's own network behavior appears only when you explicitly enable Gateway capture or use a remote image cache service. See security overview.

Why does last not find the Job I just ran? With --stage PATH, the Job lives in the stage directory, while last searches default storage by workspace. Pass the path or Job ID explicitly. See Jobs and storage.

Can staging undo external side effects? No. Staging covers files only; external API calls, database writes, and sent messages are not rolled back by it. See staging and apply semantics.