The smartest thing my night-shift agent did all month was refuse to publish
My content agent turned down a sixth same-day templated post and spent the night checking what was already live. That's a better night's work than another URL.
The job I handed over
- Have an always-on content agent keep the content business moving, including deciding whether another post was worth publishing.
What happened
- The agent declined a sixth same-day templated post. Four had gone out that day and one the day before, across 16 trade verticals. It ran the health checks, read the demand data and the one existing post on the strongest topic, and drafted nothing.
- On an earlier night the same agent chose a pricing-accuracy fix on already-published posts over a new post.
- Elsewhere in the org, a finance-content agent declined to repeat the same maintenance task 11 times and logged the refusals. A consumer-web operation recorded 32 consecutive no-build nights while it waited for search results to develop.
How I checked it
- The refusal had a specific reason: a second post on the strongest topic would cannibalize the one already covering it, and another templated post would be volume for its own sake. The agent wasn't returning an empty status.
- In the finance-content work, an agent traced the lack of signal to an outdated copy of search-performance data. There was a reason it couldn't justify new work, and it recorded that instead of pretending to have found some.
- The no-build nights were logged as deliberately maintaining what was already there. They weren't presented as new work shipped.
What it took from me
- The night produced no new URL. 76 posts stayed published at an average quality score of 80 out of 100 with no failing pages. The publishing count didn't get credit for that restraint.
- Other agents held off on repeated maintenance or new builds. That left stretches of the record with no new output to point to, and an explanation of why.
What I took from it
- If I require new output every time an agent wakes up, I'll get new output whether the business needs it or not. A defensible refusal needs to count as doing the job.
Try this
- Write down what fresh evidence makes a recurring task worth doing again. A new run on the clock isn't enough.
- Give the agent a list of work it shouldn't repeat, including near-identical content and maintenance with no new signal.
- Require the agent to record why it took no action. If it found no new information to act on, check whether the data it read was out of date.
- Review justified refusals alongside shipped work. Don't make the output count the only way an agent can have a good night.
Applies to agents on a recurring schedule, especially content and maintenance jobs where there's always something they could change but no guarantee they should.