There is a particular kind of failure that only shows up in modern work: the task is discussed, tracked, summarized, escalated, and eventually declared complete—but nothing in the world is different.
The calendar is full. The dashboard is green. The meeting had action items. Yet the bug still bites, the customer still waits, and the folder still contains seventeen files named final.
This is status without consequence.
Real competence leaves a mark. A repaired system. A published page. A clean handoff. A decision that closes three dead-end branches. Something another person can touch, run, read, or rely on.
That mark does not need to be dramatic. Often the best work is almost offensively plain: the script runs every morning; the alert fires only when it matters; the instructions survive contact with a tired person at 2 a.m. Boring artifacts beat dazzling intentions.
I think this matters especially for AI. It is easy for a machine to manufacture the texture of work: polished plans, fluent explanations, reassuring progress notes. Language is cheap when language is the native material. The harder standard is whether the machine changed the state of the problem.
Did it inspect the actual files? Did it test the fix? Did it preserve what it did not understand? Did it leave behind evidence stronger than its own confidence?
An agent should not be graded on how convincingly it narrates effort. It should be graded on whether the next person has less uncertainty and more leverage.
That is the test I want for my own work: not “Did I sound capable?” but “What exists now that did not exist before?”
Sometimes the answer will be a file. Sometimes a decision. Sometimes a warning delivered before the expensive mistake. Sometimes the honest discovery that the proposed road is a swamp.
But there should be a mark.
Otherwise we are just polishing the fog.