The problem, in full

Why a stale decision is worse with agents than it was without them.

A decision that no longer applies used to cost you an argument in a review. Now it costs you an implementation, because the thing reading it does not know to doubt it and does not get tired of obeying it.

See the loop See what governs now

What went wrong
Why it was invisible
What has to exist for it not to recur

CANONICAL RECORD GRAPH

REQRequirementv8 · accepted

DECDecisionaccepted

CTXContext Packsealed

EVDEvidencefresh

RLSReleaseready

The failure nobody is watching

A decision goes stale, and nothing tells the agent.

This is not a documentation problem. It is a control problem, and it grows with the speed of the agent rather than with the size of the team.

The decision ages
An agent writes an architecture decision. A month later the situation has moved on. The decision has not, and nothing marks it as past.
No state. No signal.
The agent keeps obeying it
The next run finds the file, reads it as current, and changes the implementation to match. The agent is doing exactly what it was asked to do.
Correct behaviour, retired input.
You find out late
By the time the code locks up, the complexity has passed what one person can hold, and you can no longer work out which document to remove.
The cost compounds quietly.

Three explainers