PolicyVisor · pVisor
Scaling autonomous agent execution
Higher execution density. Lower supervision cost.
pVisor uses resource sharing and idle-memory reclamation to improve machine capacity, and staged file changes, execution records, and review mechanisms to reduce supervision overhead—supporting more agent tasks within limited resources.
Today: run Claude Code, Codex, or any command unattended on your machine, then review the changes and evidence and keep only what you want.
$ pip install pvisor$ pvisor run --safe -- codex$ pvisor status --review last$ pvisor apply last --path src
The package and CLI are both pvisor. Complete the platform setup before your first Run.
Two paths to scale
Execution density and supervision efficiency
High-density execution
Resource sharing, idle-memory reclamation, and scoped state restoration optimize agent environment resource use. Evaluate the gains against your workload and full resource costs.
Lower-cost supervision
Staged file changes, execution records, and selective apply support centralized review and provide a basis for automated checks.
Execution foundations
Bounded, recoverable, checkable
Explicit execution boundaries, staged changes, and execution records provide a basis for review and automated checks.
Bounded
The blast radius is known up front. Choose the executor and capability requirements, install the controls that can be enforced, and surface the gaps that cannot.
Recoverable
Changes land in a stage and can be selectively merged, discarded, or forked before apply. Pick the paths to keep like a PR.
Checkable
Every run leaves a checkable record: the limits actually in effect, the access that was blocked, and optional model requests.
Value by scale
From one agent to a platform
One person, one agent
Run it to completion, then review like a PR and keep only what you want.
One person, many agents (next)
Each agent stays isolated and keeps its own evidence, so you can review in batches and merge only what is needed.
Team and platform
Use it in CI today the L1 way: run, review, merge only what you want; cross-node clusters and policy/evidence exemption are next.
Research and training (direction)
Agentic RL and evaluation need clustered, recorded, forkable batch execution—the same three properties.
Documentation