Surface — Remote tasks
Describe the task once. Come back to a diff.
Long-running agent work in a sandbox on remote infrastructure, streamed back to the browser.
- Trigger
- You, once, up front
- Execution
- Remote sandbox
- Review
- Reviewed in minutes to hours
- Gate
- Named human · on the diff
Definition
What it is
Agent runs configured up front and executed on remote infrastructure, inside a sandbox, with a provisioned workspace.
- The run streams back to the browser.
- You can intervene mid-run; mostly you do not.
Approval
How approval works
The run proceeds at full speed inside the sandbox for as long as the task takes.
- Autonomy is bounded by sandbox policy rather than by a prompt on every call.
- Anything crossing into your systems comes back as a proposal.
Fit
When you reach for it
Work that is well specified up front, long running, and not worth watching. You describe it once and review the result.
In practice
Three ways this gets used.
Engineering
Run a large systems-engineering export through a validation and normalization pass and return a discrepancy report.
Program management
Reconcile schedule data across program tools and produce a variance analysis, with the resulting updates batched for approval.
Software development
Execute a dependency upgrade across a repository — build, test, fix, iterate — on a working branch, and return the diff as a proposal.
Everywhere else
The same boundary, three other ways in.
Next
See Remote tasks configured for your tools.
A briefing walks through how a handful of your real tools would register on this surface — the classes they reach, where the gates land, and what stays identical on the other three.