Carlos Granier
Back to homelab
Agents in production · Happened

A minimum count in a prompt is a quota

My overnight agent was asked for 3 to 5 ideas and kept landing on exactly 3. A floor in a prompt turns into a target.

Fixed Accurate as of
Dot chart of ideas proposed per night by an overnight agent from September 4 to October 5. Before a prompt change on September 25 most nights land at 3 or 4, averaging 3.4. After it, nights land between 1 and 3, averaging 2.1.

The short version: my overnight agent’s prompt asked for “3–5 candidates” while its safety rules said an empty report is a correct report. Thirteen of fourteen nights came back with 3 or more, nine with exactly 3. When I changed the wording to “up to five; one or two is a normal night”, the average dropped from 3.4 to 2.1.

Get the week's seven in one email, every Monday.

What went wrong

Every night at 22:30 an agent on my server reads the state of my task records and proposes work that would be safe to run unattended. It never runs anything itself; it writes a report and I decide in the morning.

The prompt asked for “3–5 candidates”. The rules it runs under say a report with nothing in it is a correct report on a quiet night. Those two instructions pull in opposite directions, and the number won. Between September 11 and 24, 13 of 14 nights produced 3 or more candidates, and 9 produced exactly 3: the floor, every time.

What we learned

  • A floor is a target. Ask for “at least three” and you get three, whether or not three good answers exist.
  • Look for pile-ups. Clustering at exactly the number you named is the signature. Counting candidates per night made it visible in one glance.
  • The permission to return nothing has to win. If an empty answer is acceptable, say so in the same place you set the size, and say it more strongly.

What went right

The agent is report-only by design, so padding cost me reading time, not damage. And I re-read prompts when models change. That review is where the contradiction showed up.

How to do it (as of October 2026)

  1. Give a ceiling, never a floor: “up to five”.
  2. Describe a normal small result: “one or two is a normal night”.
  3. Count outcomes per run and chart them. A spike at one exact number means the prompt is setting the answer.
  4. Audit standing prompts whenever the model behind them changes.

Status

Fixed on September 25, 2026. Over the next 11 nights the agent proposed between 1 and 3 ideas, averaging 2.1. The model behind it also changed three days later, but the drop started before that. Through October 5 it never returned an empty report, so I’m still watching.

The week's seven, by email

Entries post here daily. Every Monday, one email with the past week's seven. Unsubscribe anytime.

Let's Connect