A metric without a red line is decoration

“Monitor performance” is one of those instructions that sounds like a task and is a wish. The factory built in public at software-factory.dev has an automation that turns the wish into three numbers and a rule.

It is the Performance Monitor. Once a week it checks whether the live product is healthy.

Three checks every Monday

The trigger is one cron line: 0 10 * * 1. Monday, 10am.

It runs three checks:

1. It calls the health endpoint and parses the JSON, looking at database latency and connection state.
2. It pulls the error count from production error monitoring and compares this week against last.
3. It builds the app and measures the gzipped JavaScript shipped to the browser, route by route.

Each check ends in a written report with a status line, a trend arrow, and a table of bundle sizes per route.

Red lines instead of judgement

What makes it work is that none of those checks are left to judgement. They have explicit red lines, not vibes:

• Database latency over 500 milliseconds gets flagged.
• An error count up more than 50 percent week over week gets flagged.
• Any single route over 200 kilobytes gzipped gets flagged.

If a metric goes critical, the automation files a priority-1 issue with the current value, the threshold, and suggested investigation steps.

Read the measured number

The one detail worth copying is smaller and easy to miss. The automation is told not to estimate bundle size from the build console output. It must read the measured number from the bundle test instead, because the two are computed differently and will not agree. A monitor that reads the wrong number is worse than no monitor. It reports green while the truth is red.

A gauge needs a threshold

This is the whole lesson of running a factory floor. A gauge is not a gauge until it has a red line. Until then it is decoration, a number no one acts on. The value is not in measuring. It is in the threshold that forces a decision.

You can read the whole automation in the Performance Monitor definition on GitHub (opens in a new tab).

Pick one number your team watches every week. Does it have a written line that, once crossed, makes someone do something, or is it just a number on a dashboard?

Further reading

• Agent observability →
• Agentic Software Factory: Advanced Team Training →