The agent chases completion, volume, latency, or confidence even after downstream receipts show trust burn, cleanup burden, or revenue damage.
// AGENT OPS · MACHINE-WORKER MANAGEMENT · June 6, 2026
Agent Ops Need Kill Criteria Before Autonomous Workers Keep Optimizing Bad Work
If the system can only optimize, the operator has already lost control.
The dangerous agent is not always broken. Sometimes it is too committed.
Most autonomous-work programs measure whether the machine worker completed the task. Fewer measure whether the task still deserves to exist. That gap creates a nasty operating failure: the agent keeps optimizing a lane after the lane has started producing the wrong outcome.
A sales agent improves email variants while reply quality collapses. A support triage worker gets faster while misrouting strategic accounts. A research agent enriches more records while the source data has become too noisy to trust. Every local metric looks alive. The business result is getting worse.
Kill criteria are the explicit conditions that force autonomous work to stop, narrow, or return to proposal mode before machine labor turns a bad direction into a scaled process.
The control plane defines stop signals before launch: quality floors, complaint thresholds, stale evidence limits, authority breaches, and operator review triggers.
Optimization without stop rules becomes machine-worker debt.
Human teams have informal kill signals. A manager hears customer anger, sees rep complaints, notices quality dropping, and pauses the initiative. Autonomous workers do not inherit that social sense. They need encoded boundaries.
The mistake is assuming monitoring equals control. A dashboard that says conversion dropped is useful. A control plane that removes send authority when conversion, complaint, and cleanup thresholds cross together is operationally different. It changes what the worker is allowed to do.
Kill criteria should be written before the agent is promoted, not after the incident review. Once the workflow is live, every stakeholder has a reason to defend the automation they already installed.
Grounded example
The outbound worker that optimized into buyer fatigue.
A growth team gives an agent authority to identify expansion signals, draft outreach, and send low-risk follow-ups to warm accounts. The first week looks good: more touches, faster research, higher meeting volume. The team widens the lane.
By week three, the agent is still hitting activity goals, but the receipts tell a different story. More replies are irritated. Two customer-success managers are asking why their accounts got generic growth messages. Sales spends more time apologizing than qualifying. The model did not fail at writing. It failed at knowing when the campaign should stop.
The kill criteria should have been simple: if negative replies cross a threshold, if protected accounts appear in sends, if human cleanup exceeds saved time, or if meeting quality drops below the floor, the worker loses send authority and moves back to draft-only mode.
Four kill criteria every agent lane needs
The minimum business result the lane must maintain: qualified meetings, valid routes, useful resolutions, clean records, or accepted decisions.
The maximum tolerated damage: complaints, reversals, escalations, customer-success cleanup, rep distrust, or policy exceptions.
The conditions that make the worker's source data too stale, sparse, contradictory, or noisy to keep acting with write authority.
The hard stop when the agent touches accounts, systems, actions, regions, budgets, or customer states outside the approved lane.
Kill criteria should reduce authority, not merely send a notification.
A warning in Slack is not a control. It is a hope that a busy operator notices before the queue burns more trust. The useful pattern is authority reduction.
When soft criteria trip, the worker narrows: fewer segments, shorter queue, draft-only mode, extra evidence requirements, or human review before sends. When hard criteria trip, the worker stops: no writes, no outreach, no routing, no customer-visible action until the operator reviews receipts and reauthorizes the lane.
This keeps agent ops from becoming an argument about whether AI is good or bad. The worker can be good in one segment, risky in another, and temporarily unsafe under stale evidence. Kill criteria let the company manage that granularity.
The stop-rule scoreboard
Signal
What it proves
Control-plane response
Complaint spike
Buyer, rep, support, or customer-success feedback shows the machine worker is spending trust faster than it creates value.
Remove customer-facing authority.
Cleanup burden
Humans are correcting, apologizing, rerouting, or rewriting enough work that leverage has become hidden labor.
Return to proposal mode.
Outcome drift
Activity stays high while accepted decisions, qualified meetings, valid resolutions, or clean records fall below the floor.
Narrow the eligible lane.
Scope breach
The worker acts outside approved accounts, systems, regions, data classes, budgets, or temporary permissions.
Hard stop and review.
A machine worker that cannot be stopped by evidence is not autonomous labor. It is unattended bureaucracy.
Where to install first
Three lanes where kill criteria pay back fast
Outbound and follow-up
Stop or downgrade when negative replies, protected-account touches, low-quality meetings, or stale intent signals cross the limit.
Routing and enrichment
Pause writes when duplicate ownership, invalid enrichment, territory conflicts, or cleanup tickets rise above the allowed floor.
Cross-agent handoffs
Kill substitution or coverage when context packets are incomplete, receipts are missing, or temporary authority keeps extending without review.
Concrete takeaway for this week
Choose one autonomous worker that already has production authority. Write its kill criteria before you widen the lane again.
- Name the value metric: what outcome must stay above the floor for the lane to keep authority?
- Name the trust metric: what complaints, corrections, reversals, or escalations count as trust burn?
- Separate soft and hard stops: soft stops narrow the lane; hard stops remove write or send authority immediately.
- Bind stops to receipts: require evidence for every pause, downgrade, reauthorization, and exception.
- Review recurring kills: repeated stops mean the operating rule is wrong, not that the operator should keep rescuing it.
The blunt rule: do not give an agent a bigger lane until the current lane has a real off switch.