RunAgent Sandbox

Let agents run real code without handing them the keys.

Every managed agent session gets its own isolated machine, with your credentials kept outside it. When generated code reaches for something it shouldn’t, that one action is stopped and the work carries on.

Slack@record close the books for September
Your finance appMonth-end started
ScheduleLast business day, 18:00
Own sandbox per session
close-agentMonth-end close · Finance
Reconcile 12 entitiesdone
Run generated reconcile.pyrunning
Read cloud credentials fileblocked
Upload to unknown hostblocked
Blocked in the sandboxThe run carries on
Owned issue opened · INC-221No credentials were ever inside the sandbox
What you get
Nothing shared between runs

Each session gets its own machine, isolated from every other user and run.

Enforced below the agent

Policy holds even when an agent or the code it writes does something unexpected.

No keys inside

Credentials stay outside the sandbox, out of reach of generated code.

Isolation

Its own machine for every session.

Managed agents run in a fresh, isolated machine per session, so one run can never see another’s files, processes, or data.

claims-processor · session 7712Managed
MachineIsolated · this session only
Shared with other runsNothing
Credentials insideNone
Outbound trafficChecked against policy
Its own machine, this session onlyGone when the session ends
Isolated machines

Every session starts clean and stays separate.

Nothing is shared between users or runs, and the agent never holds the credentials it uses.

Enforced below the agent

Files, processes, and network under policy.

Reading a file, starting a process, or opening a connection is checked against policy outside the agent, so it holds even when generated code misbehaves.

close-agent · generated codeSession 7712
Write recon.xlsxFileallowed
Run reconcile.pyProcessallowed
Read cloud credentialsFileblocked
Upload to paste-host.ioNetworkblocked
Stopped before it ranEnforced outside the agent
Protected execution

Allowed work proceeds. Unsafe actions never happen.

Writing to the workspace and running the job are allowed, while reading secrets or uploading to an unapproved host is stopped and becomes an owned issue.

On the record

Know exactly what the agent did on the machine.

Each file, process, and network event is recorded with the policy decision behind it and linked to the agent, the person, and the outcome.

Runtime record · session 7712close-agent
File events128
Processes started14
Network connections9
BlockedOpened INC-221 with an owner2
Every event tied to its decisionLinked to the agent, person, and outcome
Runtime evidence

Every allowed and blocked event, on one record.

The session, the policy, and the access used, linked to every event, so a security review takes minutes.

Agent Sandbox

Run one of your agents in a sandbox.

Pick an agent and a real task. We’ll run it in a sandbox and show what’s allowed, what’s stopped, and what’s recorded.