That night taught the first and most costly lesson in operating an agentic system: the most dangerous failures are not loud crashes; they are silent absences. A model can hallucinate, a prompt can go sideways, a secret can leak — these are the disasters that appear in pitch-deck nightmares and trigger pages at three in the morning. But the routine that exited early because its guard asked the wrong question? Nobody noticed. The queue that stayed empty because the eligibility criteria broke? Still quiet. A scheduled job that produced nothing for three weeks, while the system watched in silence? This is the shape of production failure in agentic systems: not a bang, but the absence of work that should have happened.
People discover most serious problems in agentic systems by noticing an absence, not a failure. The cure is discipline: a habit of checking for nothing.
Start by reading absence on a cadence — not when something goes wrong, but regularly, every week. Open the routine catalogue and scan the time series of recent runs. Look for routines that have produced nothing recently. This is not inherently bad. A routine that searches for a specific condition and finds it absent is behaving correctly. But if the routine used to find things, and now finds nothing, and you cannot immediately name why, the quiet is suspicious. Fourteen straight empty runs is a story, even if the audit log does not say "failure." Empty runs are still runs. The log records them alongside meaningful ones. A time series of emptiness is data. Read it.
When you see a routine that has gone quiet, cross-check the story against your tracker. Your tracker holds the source of truth: which records were eligible for this routine to act upon? If the tracker says five items were eligible and the routine found nothing, the mismatch is the bug. Usually, the mismatch lives in one of two places — the tracker mapping is wrong, or the routine's eligibility query does not match the tracker's definition of eligible. Rarely is it the model. Start with the plumbing before you start with the magic.
Make the habit stick with one small discipline before you open the workspace each week. Ask yourself a single, honest question: what should have happened this week? Don't answer it. Just name it. Write it down in a sentence. "The crawler should have found three new sources." "The extractor should have clarified at least one ambiguous claim." "The routine should have surfaced one routine-maintenance task." Naming the expectation makes silence visible. When you open the audit log and find zero runs, or three empty runs, that expectation sits in your mind as a mismatch. The system is not lying — it is answering a question you never asked it to check.
This is the critical move: treat sudden quiet like sudden noise. A routine that stops producing anything after two weeks of steady work deserves the same investigation as a routine that suddenly emits ten findings a day. Both are deviations. Both break the pattern you learned to expect. A system that does nothing unexpectedly is as broken as a system that does everything unexpectedly.
Cultivating the habit#
The hardest part is not technical. People are wired to react to alarms — a page, a red dashboard, an error message. Nobody naturally schedules a recurring task to look for nothing. The cure is institutional: a small, dull, weekly ritual owned by someone on the team. Fifteen minutes, last thing on Friday, before the weekend starts. Open the audit log. Filter on each routine that matters. Look at the time series of runs and results. Cross-check against your tracker if anything went quiet. Close the browser. Done.
This is not a meeting. This is not a committee. This is one person, one Friday afternoon, one spreadsheet of expectations, armed with a habit. The cost is real — fifteen minutes a week — but the avoided cost is the Monday morning when a customer notices something Ship was supposed to do. The avoided cost is shipping nothing on a Friday and nobody knowing until the following week when someone external complained.
A quiet workspace is healthy when the quiet has been checked. A quiet workspace where nobody has looked for absence is the same as a green dashboard with a broken alerting pipeline — you have no information about the shape of the system. You have only the absence of noise. Open the audit log every week. Look for nothing. Name what should have happened. Make the mismatch visible. That is the discipline.