A score created before a funding event, hiring freeze, pricing change, or territory move should not keep routing work unchanged.
// AGENT OPS · GOVERNED AGENT LABOR · June 6, 2026
Agent Ops Need Expiry Dates Before Machine Workers Run Old Judgment Forever
Autonomy without expiry becomes bureaucracy at machine speed. Every meaningful agent judgment needs a time limit.
A one-time human yes can quietly become permanent authority when agents reuse it as precedent.
Do-not-touch states protect trust until they become forgotten revenue black holes.
Machine workers can keep assigning work to teams whose capacity, coverage, or accountability already moved.
The oldest decision in the system is often the most dangerous one.
Most agent-ops reviews focus on the newest action: what the worker did today, what it wrote, what it changed, what it escalated. That is useful, but it misses the quiet failure mode of AI-native operations: agents keep executing old judgments as if time has no cost.
A worker scored a prospect as low priority three weeks ago. Since then the company raised a round, hired a VP Sales, and started searching for tools in public job posts. The agent still suppresses outreach because the old score has no expiry. Another worker paused a renewal account during a support incident. The issue closed, the champion returned, and the account is still suppressed because nobody forced the suppression to refresh.
That is not intelligence. That is stale context wearing an automation badge.
Machine workers do not naturally know when old truth goes bad.
Human operators often carry time in their heads. They know last quarter's territory model is obsolete, last month's pricing exception was temporary, and last week's customer hold should be reviewed after the executive call. Agents need that operational memory encoded.
Without expiry dates, an autonomous worker can compound decisions long after the reason disappeared. It can keep leads out of queue, keep accounts in cooldown, keep discounts blocked, keep tasks routed to the wrong manager, and keep forecasts shaped by evidence that no serious operator would still trust.
The problem is not that agents are careless. The problem is that the operating system treats a decision as valid until someone notices it is not. At machine scale, “until someone notices” is a bad governance model.
Four judgments that need shelf life
Fit, intent, churn, expansion, support risk, forecast confidence, enrichment quality, and propensity scores should carry a generated-at time, data sources, and refresh trigger.
Human approvals should state whether they are one-time, case-limited, account-limited, campaign-limited, or reusable precedent.
Quiet periods, consent windows, legal holds, customer pauses, low-quality filters, and capacity blocks need review dates before they become invisible prisons.
Owner assignment, territory mapping, escalation paths, SLA lanes, and reviewer queues should expire when org structure, load, or accountability changes.
Grounded example
The account that stayed buried after it became valuable.
A growth team uses a machine worker to decide which accounts deserve founder-led outreach. In March, the worker suppresses a mid-market account: weak hiring signal, no recent funding, low product fit, and a prior unsubscribe from one contact. Reasonable call.
In April, the account launches a new region, posts five revops roles, and installs two adjacent tools. The worker sees the new data, but the suppression rule outranks fresh intent. The original block has no expiry date and no review owner. The account remains buried while a competitor wins the buying cycle.
The fix is not “let agents ignore suppression.” That would trash trust. The fix is a better contract: suppress for 30 days, reopen only on material account change, require a new contact path, and route to a human if consent is ambiguous. The agent does not need bravery. It needs a shelf-life policy.
A decision expiry record should answer this.
Decision
What judgment the worker is allowed to reuse: score, approval, suppression, route, owner, risk label, exception, or customer state.
Born from
The evidence, policy, owner, model version, timestamp, and business context that made the decision valid.
Expires when
A fixed date, time window, data age threshold, context shift, owner change, policy version change, or downstream capacity change.
Refresh path
What the worker must do when expiry hits: rescore, ask, sample, escalate, quarantine, renew approval, or stop execution.
Extension owner
The accountable human or policy owner allowed to extend the decision, and the evidence required for extension.
Expiry does not slow autonomy. It prevents fake autonomy.
Bad teams hear expiry dates and imagine more review meetings. Good teams hear expiry dates and remove thousands of hidden decisions from human memory. The worker can refresh routine judgments automatically, escalate only material changes, and stop only when the underlying business fact has aged beyond tolerance.
That is operator leverage. Instead of asking a manager to remember every temporary exception, the control plane carries the timer. Instead of reviewing every agent action, leaders review decisions that survived too long, refreshed too often, or expired without a clean owner.
Machine labor becomes faster when stale judgment is impossible by design.
If an agent can reuse a decision forever, you have not delegated work. You have automated a ghost policy.
Operator audit
Find stale decisions this week
Oldest active score
Pull the oldest scores still affecting routing, suppression, prioritization, or customer treatment. If nobody trusts them, expire them.
Longest suppression
Find accounts, leads, tickets, or renewals blocked for more than one cycle. Confirm whether the reason still exists.
Reusable approval
Review approvals agents treat as precedent. Mark which ones are one-time, limited, reusable, or revoked.
Concrete takeaway for this week
Add expiry to one high-impact machine-worker decision before increasing scope.
- Pick the decision: lead score, account suppression, approval, route, or customer hold.
- Set the shelf life: define max age, context-change triggers, and evidence required for refresh.
- Encode the stop: when expiry hits, the worker must refresh, escalate, quarantine, or stop — not keep going silently.
- Review exceptions: stale decisions that get extended repeatedly are signals of broken policy or missing ownership.
Autonomous companies do not just need fast workers. They need workers that know when yesterday's judgment has expired.