Define the error rate, rollback rate, or bad-output pattern that immediately pauses agent authority.
// GOVERNED AGENT LABOR · MACHINE-WORKER MANAGEMENT · June 6, 2026
Machine Workers Need Kill Criteria Before Autonomy Keeps Failing Forward
If an agent can keep failing forward, it has authority without management.
Autonomy needs a stop condition.
Most agent programs define how machine workers start work. Far fewer define exactly when they must stop. That gap is where bad autonomy becomes expensive. The agent misses a signal, retries with a slightly different plan, asks for another exception, patches the symptom, and keeps moving. Each step looks reasonable in isolation. Together, they turn failure into operating policy.
Kill criteria are not drama. They are management infrastructure. They say: when this threshold is crossed, the worker loses authority, the queue freezes, and an operator decides whether to repair, retrain, constrain, or retire the agent.
The hard part is cultural. Teams want autonomous agents to feel productive, so they treat stopping as a failure. It is the opposite. A machine worker that knows when to stop is safer, more promotable, and easier to trust.
Stop work when human review, exception cleanup, or customer-facing repair exceeds the value created.
Freeze autonomy when the worker starts changing behavior outside approved boundaries or source-of-truth rules.
Require receipts before restart: affected scope, cause, rollback path, and a changed control.
The bad pattern is endless remediation.
The agent keeps negotiating
The worker fails a task, opens a follow-up, changes a prompt, asks for a broader tool permission, and continues touching production records while everyone calls it iteration.
Authority drops on contact
The worker crosses a defined threshold, loses write access, packages evidence, and waits for an operator decision. No more retries until the control plane changes.
This matters because machine workers do not fail like humans. A person usually slows down when confused. An agent often accelerates: more calls, more tickets, more edits, more confident summaries. Without kill criteria, the system rewards persistence even when persistence is the dangerous behavior.
Ground example
Campaign agents can damage trust before metrics show it.
Take a GTM agent managing outbound campaign hygiene. It refreshes account fields, selects personas, updates sequence membership, and drafts follow-up variants. For a week, the numbers look strong: more accounts processed, more contacts enrolled, fewer stale tasks.
Then sales starts complaining. Enterprise accounts receive SMB copy. Existing customers get prospecting messages. One region's territory rules are ignored because the enrichment source conflicts with CRM ownership. The agent keeps producing activity while humans quietly clean the mess.
A kill-criteria system would stop this early. Three customer/prospect mixups in a day freezes enrollment authority. A rollback rate above 5% pauses CRM writes. Any conflict with account ownership moves the worker into recommendation-only mode. Restart requires a receipt: what changed, which records were touched, which source won, and why the same failure cannot repeat.
Three rules make kill criteria operational.
Do not invent the stop line after the agent is already politically useful. Predefine error, cost, drift, customer-risk, and review-load limits.
Stopping one job is weak if the worker can keep writing through another route. Kill criteria must bind tools, queues, permissions, and automations.
No vibes-based relaunch. Require cause, affected scope, rollback status, owner decision, and a concrete guardrail update.
The goal is not to punish agents. The goal is to make autonomy legible. Humans trust systems that can stop themselves under known conditions. They distrust systems that keep explaining why the last mistake was special.
A machine worker without kill criteria is not autonomous. It is unsupervised persistence.
Concrete takeaway for this week
Pick one production or near-production agent and write its stop conditions before expanding scope.
- Set the lines: error rate, rollback rate, review minutes, customer exposure, data drift, and policy conflict.
- Bind authority: decide which tools, queues, write permissions, and automations freeze when a line is crossed.
- Define the receipt: affected records, root cause, rollback status, owner decision, and changed control.
- Practice once: run a small recall drill so operators know how the worker stops before it matters.
The blunt rule: if you cannot say when an agent must lose authority, you have not governed it. You have only hoped it behaves.