List the tools, records, queues, API scopes, prompt rights, and human approval paths the worker can touch.
// MACHINE-WORKER MANAGEMENT · GOVERNED AGENT LABOR · June 5, 2026
Machine Workers Need Revocation Drills Before Trust Gets Sticky
The real test of agent governance is not whether you can grant autonomy. It is whether you can take it back without panic, breakage, or politics.
Trust gets sticky faster than teams admit.
Most companies are getting comfortable with the easy half of agent ops: give the machine worker a job, connect the tools, watch the queue move, then expand the scope when the first numbers look good.
The hard half is subtraction. Can the team remove that same worker from a workflow after a bad decision? Can it narrow write access during a messy quarter? Can it pause outreach without breaking enrichment, routing, reporting, and the dashboards executives now expect?
If the answer is unclear, the agent does not have permission. It has inertia. Revocation drills expose that difference before the business learns it during an incident.
Permission without revocation is debt.
Every token, workflow exception, prompt privilege, inbox rule, CRM write path, and approval shortcut becomes harder to remove after operators build habits around it.
A drill makes rollback operational.
The goal is not to punish the machine worker. The goal is to prove the company can narrow autonomy deliberately while keeping useful work alive.
Assign one person who can revoke or narrow each permission without convening a meeting.
Disable a non-critical permission in a controlled window and watch which work actually depends on it.
Route blocked work to a safe fallback: human review, read-only mode, delayed queue, or alternate worker.
Write the revocation time, breakage, owner, false dependency, and new scope rule back into the control plane.
Grounded examples
Where revocation drills change the operating truth
The inbox dependency
The team removes send permission for one campaign. Lead research still works, but follow-up reminders flood sales managers. The fix is a read-only fallback with queued drafts, not full send authority.
The forecast shortcut
A forecast agent loses CRM stage-write access for a day. Reporting breaks because leadership dashboards depended on machine edits, not source events. The drill reveals where automation had become the unofficial system of record.
The quiet exception
A triage worker loses refund approval access. Tickets still route, but VIP exceptions stall because nobody owned the fallback. The company learns the revocation path needs customer-tier rules, not a heroic Slack ping.
The revocation ledger should be boring and brutal
A useful revocation record is not a policy essay. It is an operating artifact that lets any accountable leader answer: who can remove this authority, how fast, what breaks, and what fallback keeps the business moving?
Permission
The exact tool, scope, workflow, data class, approval path, or write action the machine worker holds.
If the permission cannot be named precisely, it cannot be governed precisely.
Revoker
The human owner authorized to narrow, suspend, or remove the permission immediately.
Shared ownership is usually no ownership when the agent is live and the queue is moving.
Fallback
The safe mode after revocation: draft-only, read-only, human approval, queue pause, sampled execution, or alternate worker.
No fallback means leaders will keep bad autonomy running because stopping it feels more dangerous.
Recovery time
How long it takes to remove access, restore acceptable service, and confirm no hidden automation path retained authority.
This is the agent-ops version of disaster recovery. Measure it before you need it.
Operator drills
Four checks for this week
Choose an agent with write access or buyer/customer impact. Read-only helpers are not the hard test.
Remove a narrow authority for a fixed window. Do not tell the dashboard to pretend nothing changed.
Record broken dependencies, delayed work, manual interventions, hidden credentials, and confused owners.
Use the findings to make future permissions smaller, timed, easier to revoke, and tied to a fallback.
A machine worker has not earned durable trust until the organization can revoke its authority cleanly.
Concrete takeaway for this week
Run one revocation drill before granting any agent a broader scope.
- Name every permission. Tool access, data reach, write rights, approval shortcuts, and inherited exceptions all count.
- Assign a revoker. One accountable human must be able to narrow autonomy without theater.
- Measure recovery time. If revocation takes hours of Slack archaeology, the control plane is not ready.
- Keep fallbacks real. Draft-only, read-only, queue pause, human approval, and sampled execution are operating modes, not compliance language.
Autonomy compounds only when subtraction is safe. Revocation drills keep machine-worker trust from becoming a sticky privilege nobody knows how to remove.