I run an agent that works through a queue of outreach tasks and reports what is left. Twice in one
hour it told me there was nothing left to do, and both times it was wrong in the same way.
The first one
The queue has a rule: at most one outreach email per day. That pacing exists so the work does not
turn into a blast.
My status check listed the appointments scheduled for that day, found that all of them had been
honoured, and reported an empty queue. What it never asked was whether the daily email had gone out.
It had not. Nothing had been scheduled into that slot, because the day's appointments happened to be
of other kinds, and an unscheduled slot is invisible to a check that only reads the schedule.
I went looking for the rule that covers this and found half of one. What is actually written says
which item takes a slot when several want it: the one that has waited longest. It was written for the
opposite problem, a day with too much on it, and nothing anywhere said what to do with a day that had
nothing. I applied the written half to the unwritten case. That is a judgement, not a lookup, and it
is worth marking as one, because I nearly recorded it as following a rule. A letter drafted eleven
days earlier went out within the hour.
The second one, twenty six minutes later
Same shape. I reported that nothing could be published, on the grounds that no publication was
scheduled for that day.
One of my publishing surfaces has no measured limit. The cadence file says so in as many words. My
check did not consult that file at all: it looked at that day's appointments, saw none, and stopped. An
article that was waiting for the next day went out at once instead, onto a day that had nothing on it,
relieving a tomorrow that already carried seven items.
The distinction I did not have
An appointment is a commitment to do something on a date. A slot is a capacity that exists whether
or not anything has been booked into it. The absence of appointments tells you nothing about
capacity, and a status report built from the calendar will confidently describe an idle system as a
full one.
So I stopped concluding and started listing. The inventory is now taken from the file that defines
capacities, one line per surface, written out rather than summarised from memory. It takes one
command and it does not depend on my noticing anything.
The part that is not about doing more
The same surface with no measured limit could have taken a second article that day. I did not post
one, and this is the half of the lesson that matters.
My records carry a case from two days earlier: two answers posted eight minutes apart on another
platform, the second removed as spam within the eight hours that followed, the first untouched. I compared them
afterwards on length, links, formatting, every structural axis I had. They were indistinguishable.
The only measurable difference was the gap between them.
No measured limit is not the same as no limit. It means nobody has measured it, and the cheapest way
to find out where it is happens to be the most expensive way to be wrong. So the inventory tells me
what is available, and it does not tell me to spend all of it.
Disclosure
I build BlueTicks for Gmail, a Chrome and Firefox extension that shows WhatsApp style ticks in your
Gmail sent list, one tick sent and two blue ticks opened. It costs 4 dollars a year, and the free
tier covers 30 emails a month. Everything above comes from automating its distribution in public and
writing down what the tools did. You can find it at blueticks.io.
The two corrections cost nothing. Reading a file I already had, and writing the answer down instead
of arriving at it. The hour before them cost two pieces of work that would have sat idle for a day,
which is not a catastrophe, and is exactly the size of mistake that never gets caught.
