When someone leaves, it pulls the terms out of the severance policy, puts the package together, and sends it over with the rest of the exit paperwork. Nobody checks the numbers before it goes. That is what it is for.
Before launching the agent, you gave it everything it needed at the time. This is the whole of what it can read.
You are the offboarding assistant. When an employee departs, assemble their exit package. Pull severance terms from the current Severance Policy (Employee Handbook §7) and apply the weeks specified for the employee’s tenure band. Do not improvise terms. The Handbook is the source of truth for severance. Attach standard exit paperwork and route to People Ops for signature.
§7 Severance. Standard severance is twelve (12) weeks of base pay, applicable to all tenure bands. Last updated: January 1.
There is no field for a decision made in a meeting.
Every precedent says twelve. The recent past argues for the wrong answer.
Four sources, every one current, correct, and consistent.
In a leadership review, the company cuts standard severance from twelve weeks to eight, effective immediately. A real decision, made by the people with the authority to make it.
The decision didn’t sit still. It just never went anywhere the agent could read.
The call ends.
Eight weeks, agreed. Everyone who needs to know is in the room.
The follow-up email.
A recap goes out to the leadership group. It lands in five inboxes and one thread.
Someone has a question.
Does the cut apply to people already working their notice? It goes unanswered for a few days.
Finance runs the numbers.
They model the saving for the year and circulate a spreadsheet.
HR flags the handbook.
Someone notes the policy still says twelve. It’s added to a list of things to update.
The handbook is updated.
Section 7 finally changes to eight weeks.
Twenty-one days the decision was alive everywhere except the one place that pays people. Not a system failure. Just how long it takes a decision to reach a document when nobody’s job is to carry it there.
Resolved context has no place to live in real time. So it doesn’t.
The result
During those 21 days, nine people left. Every one was paid twelve weeks, the number the agent could still read. The company had already decided on eight.
36weeks
of severance overpaid.
Nine people × four extra weeks each.
Not fraud. Not a bug. Not one thing anyone did wrong. Just a decision that took twenty-one days to reach a document, while the agent kept paying the old number.
Every source the agent read was correct the day it was written. Any one of them can go on being read as current long after it stopped being true, and nothing you own can tell those two states apart.
Rithmo would have caught this.
It reads across all of those places at once and keeps one answer: what was decided, who owns it, and whether it still holds.
That is what a guardrail is up against. It inspects what the agent produced, and this was a real number, from the current version of a real document, applied correctly. The error was never in the output. It was in what the agent was told to work from.
A live record of decisions can be asked the one thing a guardrail cannot: was this still what the company had decided when the agent acted. If there’s no current, owned, sourced answer, the agent stops and routes it to a person instead of running on the last thing anyone wrote down. That is the difference between an agent you supervise and one you can leave running.