Background engineering work

Give agents bounded maintenance tasks that finish as reviewable changes.

Connect customer-configured compute for repository files, shell, git, and tests. Dependency updates, failed CI investigations, small fixes, and maintenance can then finish as a branch, PR, transcript, and activity record.

Identity and mediated activity

Keep the caller, its grants, and tool-call results inspectable.

Each named Agent identity keeps its authority and mediated tool-call records together. The activity log does not claim visibility into commands or credentials used directly inside a connected execution environment.

Maintenance Agent identity with a scoped GitHub grant
The named caller carries its own grant scope, mode, and lifecycle.
Recent mediated GitHub activity showing successful, failed, and approval-paused calls
The identity view keeps several outcomes visible without presenting local machine work as gateway activity.

Start here

Give maintenance work a clear finish line.

Start with maintenance work that finishes as something reviewable: a report, a branch, a PR, test output, and a record of what the agent did.

Small task

Start with one issue or maintenance job.

A narrow objective is easier to scope, easier to review, and easier to repeat later.

Review output

Prefer branches and PRs over direct changes.

The agent can produce a diff and test output without needing merge authority.

Repeat later

Promote only trusted paths.

After the task works manually, schedules, events, or workflow runs can reuse the same scoped controls.

Examples

Maintenance jobs that finish in review.

1Dependency update with test output and a branch.
2Failed CI investigation that proposes a focused fix.
3Repository maintenance with activity records for review.

Get started during beta

Start with one bounded task and one boundary to review.

Use the app during beta, or open the docs and choose the first agent task you want to govern.

Start an agentRead GitHub quickstart