Agent Managers Need Exception Half-Lives Before Queues Become Graveyards
Most agent managers do not drown in volume. They drown in stale exceptions. A machine flags a weird lead, a pricing mismatch, a routing conflict, a support edge case, or a risky outbound draft. The queue fills. Nobody decides how long those exceptions may live before the system escalates, expires, or freezes dependent work. Soon the queue is not governance. It is a graveyard with status labels.
Queues keep every anomaly forever, so urgent risk, useful review, and dead noise all compete for the same human attention.
Exception half-lives assign a timer and a consequence: escalate, auto-expire, block downstream work, or force model retraining.
Operators regain trust in the queue because every unresolved exception is moving toward a deliberate outcome instead of rotting in place.
What queue rot looks like in the wild
The quote agent keeps flagging non-standard discount stacks, but sales keeps sending manual workarounds while the exception queue never resolves the underlying rule.
Territory disputes sit open long enough that reps create off-system side agreements. The queue becomes historical fiction.
The renewal agent keeps seeing unresolved product incidents, but the exception ages past the account decision window and the machine nudges expansion anyway.
What the half-life actually decides
How long can a named operator ignore the case before it jumps a level?
When does the evidence go stale enough that the queue item must die instead of haunting future decisions?
At what age does the open exception block machine writes tied to pricing, routing, messaging, or account state?
When repeated exceptions on the same pattern force rule repair or model tuning instead of more human cleanup?
The exception half-life ladder
Different queue items deserve different decay curves. Treating every exception as “review someday” is the lazy path to fake governance.
Concrete operating examples
Queues are only governance if age changes consequences.
What good looks like
Humans see the live edge cases, not a landfill of ancient machine uncertainty.
High-risk exceptions can halt downstream automation before one stale case infects live revenue work.
Recurring queue patterns become product, policy, or prompt fixes instead of permanent human chores.
The common failure pattern
Teams celebrate that they “have human review” because every weird case lands in a queue. Then they never define aging rules. The machine keeps generating exceptions faster than anyone resolves them. Operators start bulk-clearing noise, sales invents unofficial workarounds, and leadership concludes the agents are unreliable. The agents were not the main problem. The queue was. It had no physics.
Concrete takeaway
- Tag every exception class by business consequence: revenue-critical, customer-risk, data conflict, or learning debt.
- Assign a half-life to each class this week, not next quarter.
- Decide what happens when the timer hits: escalate, expire, freeze, or trigger rule repair.
- Make exception age visible in the operator dashboard so the queue cannot pretend all items are equal.
- Review recurring exceptions weekly and kill the pattern at the rule layer instead of funding more cleanup labor.