For eleven nights, an agent I had set up produced a "dream": a narrative synthesis of the previous 72 hours of activity. Then nothing. The system didn't crash. It shut down silently, and I didn't notice it had stopped for 46 days. It's the most instructive failure of my year, because the mechanism isn't limited to a personal workspace: it threatens any AI project that has no reader.
The experiment
An agent ran every night at 3 a.m. through a scheduled cloud routine. Its task: read the activity of the last 24 to 72 hours (session logs, commits, open threads) and turn it into a "dream". Not bullet points: narrative prose, meant to surface the underlying thread that no dashboard shows. Fourteen dreams were generated between April 20 and May 22, 2026, 5 to 7 KB each.
What broke, and it wasn't the expected cause
The routine ran on cloud infrastructure, supposed to execute around the clock without depending on any machine. Yet it had a hidden dependency: it expected a state that only existed on the local computer. The PC is switched off at night, so the "cloud" routine broke as early as May 1. Eleven days later, I found out while auditing the setup. I rebuilt it to run 100% locally: the dreams came back, then, on May 22, total silence. For 46 days, no output. I didn't notice.
A so-called "autonomous" routine that secretly depends on a local state isn't autonomous. It's living on borrowed time.
Why nobody saw it
Two causes, and the second is the real one.
- Presence ≠ execution. The setup existed, documented, visible on the dashboard. But nothing checked that it actually ran every night. A system that is present is not a system that executes.
- No consumer. The dreams had a producer, the agent, but no confirmed reader. Nothing required anyone to read them. An artifact nobody consumes isn't missed when it disappears. So it disappears.
A system with no verified trigger AND no consumer drifts and dies silently. It isn't a failure, it's an extinction.
What this says about your AI projects
An AI proof of concept rarely fails because the model is too weak. It fails because:
- nobody wired a reliable trigger: it runs "when someone remembers";
- nobody checks that it actually executes: presence is taken for execution;
- nobody consumes the output in a real decision: it's clean, but orphaned.
A "dream" dying out in a personal workspace is anecdotal. The same mechanism on a support assistant, a scoring model or an automated watch feed means a burned budget and lost trust, without a single error message.
The rule to draw from it
Before putting an AI system into production, three questions, in order:
- Trigger: what makes it run, and how do you know it ran?
- Execution: what signal proves it executed, and not just that it exists?
- Consumer: who reads the output, and in which decision? If the answer is "nobody", it shouldn't have been built.
A system with no consumer doesn't need fixing. It shouldn't have been built.