Use cases¶

The same pVisor solves different problems at different scales.

One person, one agent¶

Let Claude Code or Codex complete a refactor fully unattended: it edits files and runs tests, and may delete a few things along the way. You do not watch; when it finishes you review the changes as you would a PR and merge only the paths you want. Available today.

pvisor run --safe -- claude
pvisor status --review last
pvisor apply last --path src

One person, multiple agents¶

Run several agents on different approaches at once. Each Job is isolated and leaves its own evidence, so you can review in batches and merge only the results you need. Next (L2); see parallel agents.

Teams and platforms¶

Put pVisor in CI: let the agent finish, review, and merge only what you want. Available today (the L1 way); policy- and evidence-driven exemption and clustering (L2/L3) are the direction—see running agents in CI.

Research and training¶

Agentic RL and evaluation want exactly a "large-scale, untrusted, recordable" execution substrate: trajectories can be recorded, forked from checkpoints, and replayed by tool prefix. Direction; see RL rollouts and research directions.