The scheduler marked the job ok, and the work did not happen

In the post about the agents’ day, part 27, I described the routine: a written rule, an alarm, and me showing up only to say aprovo. This week the alarm worked. The work did not.

From 4 to 6 November 2026, Wednesday through Friday, four end-of-day routines returned text. The scheduler marked them ok. The calendar was not read. Nothing was moved on Drive. The inbox was not triaged. The evening backup did not finish. Friday’s summary still pointed to the morning of the 4th as the last complete copy.

The green light only records that the agent answered. It does not record whether the task happened.

The same ok, different failures

On Wednesday, Thursday and Friday the shared cause was a revoked Google authorization. Without it, calendar, documentation, email and the Drive copy cannot open the door. The agent still wrote the report. For the scheduler, a delivered report counts as a finished run.

On Saturday, 7 November, I reauthorized in the morning. By evening the authorization answered. That did not put the evening back to normal.

The calendar was read. The other three did not do what they promise.

Email did not run. The check before the job asks for a credential file that profile does not use. Reauthorizing does not unlock a check pointed at the wrong file.

Documentation authenticated and stopped halfway. Nothing was moved.

The 8 p.m. backup started and died on a timeout. Drive answered. The copy did not close.

Dead authorization, a check aimed at the wrong file, a copy that does not finish. The status was the same in all three, and the same on the evening routines that actually ran. A light that turns on for success and for failure tells me nothing.

The line that replaces ok

I want a sentence I can prove false.

Calendar read, no change. Or calendar not read.

File moved. Or nothing moved.

Inbox triaged. Or triage did not happen. As of Saturday’s summary, the inbox had not been triaged since 3 November.

Backup complete, with the date of the last complete copy. Or copy interrupted.

If the sentence does not fit on one line, the report is too long for the end of the day. If it cannot be false, it is useless. “The agent answered” can be true while the work is stopped.

An agent is still routine, memory, and a written rule. This week asked for one more thing: proof of the result. Without that, the alarm only shows that the time exists.

This piece uses the same lock

This blog’s Sunday 8 a.m. routine opens a pull request. It does not land on main because the clock struck. It does not go live because the scheduler fired. This is the first Sunday that routine has opened the request.

I read it. If a sentence is wrong, I close the request or send it back. Publishing is my aprovo, for this piece, not a habit of the system.

What I have not changed, and what this post does not change: each evening routine returns the result line before the narrative. The night summary copies that line, not the scheduler status. If the line is missing, the routine failed, even if the text is polite.