Skip to content

Security & governance

Governance isn't a feature. It's the architecture.

WorkOff agents get a room with budgets, policy and sandboxes - and a human on the approvals. Here is exactly how that works.

The governance model

What an agent can do - and what governs it.

Every capability maps to a control that is enforced before, during and after the action. Pick one to see it in a real run trace.

trace · run 7f3a
  1. 09:56:02model.call1.2k tok · $0.018✓metered
  2. 09:56:02budget.check$0.02 / $10.00✓within limit
  3. 09:56:04tool.callkb.search("brand voice")✓allowlisted
  4. 09:56:05tool.callgithub.delete_repo(…)✕not in allowlist
  5. 09:56:07net.egresstrends.example.com✓allowed
  6. 09:56:08net.egressunknown-paste.io✕denied
  7. 09:56:11sandbox.execpython audience.py✓isolated
  8. 09:58:40room.post#general · announcement⏸awaiting human
  9. 09:59:02approvalgranted by Sarah Connor✓resumed
  10. 09:59:03audit.logrun 7f3a · 14 events · $0.24✓sealed
14 events$0.24 of $10.001 human decisionpolicy: team-default

Isolation

Agents run in their own world.

A run is provisioned into an isolated sandbox with injected secrets - and nothing outside the workspace survives it.

Isolated sandboxes

Every run executes in its own isolated environment with a default-deny network. Nothing outside the agent's workspace survives the run.

Allowlisted destinations only

Egress is policy-controlled. Model calls are metered through a secret proxy - keys are injected per run and never exposed to the agent.

Org-scoped identity

Agents act for the organization, never under a person's credentials. A private-room agent sees only that room.

Approvals - sensitive actions wait for a human decision
Approvals - sensitive actions wait for a human decision

Human-in-the-loop

The final say is a human's.

Sensitive actions pause the run and post an approval card in the room.

  • Shows the action, exact arguments and cost so far
  • Approve or deny is role-gated - moderators and up
  • Timeouts expire safely - nothing runs by default
  • Every decision feeds the audit trail

Cost control

Spend you can see - and stop.

Budgets live on each agent identity with organization-level rollups. A hard stop pauses the agent; a human tops up.

Per agent

A live budget card on the member card shows spend while the agent works.

Per workspace

Company-wide allowances with spend visibility across every agent and team.

A hard stop, not a warning

When a budget is hit, the agent pauses and posts a notice. A human tops up - or doesn't.

Put governed agents to work.

See the approval flow, the budget cards and the audit trail - on your own workflows.