Three operating systems
Windows, Linux, and macOS sandboxes from one API, for agents that operate real desktop and server software.
Recorded with the runImage and OS attestation
Infrastructure
Agent Sandboxes
Own the machines your agents run on.
The sandbox service behind Engine, available standalone. Deploy it on-prem or in your cloud and give every agent an isolated Windows, Linux, or macOS machine with snapshots, network policy, and full session recording.
Agent Sandboxes
Example workflow
Running
Linux
us-west / private
Ready
Windows
on-prem / finance
Ready
macOS
build pool / 04
Network policyrefund-review
api.internalAllow443
documents.internalAllow443
Public internetDenyall
Checkpoints
Image ready
snapshot 01
Tools installed
snapshot 02
Before submission
snapshot 03
How it works
From request to a reviewable result.
Each run keeps the request, actions, evidence, and final decision together.
01 · Create
Start from an approved image.
Request a Windows, Linux, or macOS environment with the machine policy assigned to the workflow.
Output · Isolated machine
02 · Operate
Run with scoped network access.
The agent uses the machine while Context records actions and enforces the configured egress policy.
Output · Recorded session
03 · Preserve
Keep evidence, not residue.
Export artifacts and the run record, then restore a checkpoint or destroy the environment.
Output · Artifacts and snapshot
What you can inspect
Capability is only useful when it leaves evidence.
Your infrastructure
Runs on-prem or in your VPC. Machine images, data, and traffic never leave your perimeter.
Recorded with the runDeployment and region
Snapshot and restore
Checkpoint a machine mid-task, branch it, or roll it back. Reproduce any run from its snapshot.
Recorded with the runImmutable snapshot ID
Network policy per sandbox
Default-deny egress with allowlists per workflow, so an agent reaches exactly what its task requires.
Recorded with the runApplied egress policy
Session recording
Full video and input capture of every session, tied to the run's trace for review and audit.
Recorded with the runTime-linked recording
Deployment boundaries
Control is part of the product.
Agent Sandboxes can run as a standalone surface, while its traces, evals, and policy decisions remain part of the same Context system.
Capacity follows deployment
Operating-system availability and startup time depend on the images and hardware configured in your environment.
Recording is configurable
Retention, redaction, and access to session recordings follow your organization's policy.