Work record / Stories

The machine I thought was idle was the one on fire

An audit showed my picture of which subscription plan carried which machine was backwards.

My own work, plus a growing fleet of scheduled AI tasks, runs over three machines and two separate subscription plans. I carried a firm mental picture of which plan bore which machine's load. An audit showed that picture was precisely backwards.

I quit sorting automation by the physical machine hosting it and sorted by function. One plan carries my interactive work and nothing else. The other carries every scheduled and background task, all moved to the single computer that never shuts off.

The plan I believed was burning was quietly handling both my live work and the whole nightly fleet. An idle machine costs close to nothing. What costs money is running the wrong kind of work on the wrong host. I also ruled that an expensive, careful model tier belongs only on my live work or true synthesis. It never runs unattended in the background, because a scheduled task's error compounds while no one watches.

The routine fleet now runs on cheaper, faster models on the always-on machine, and my interactive work keeps the stronger model. The fleet writes its own status file so each machine sees one shared picture, in place of every machine guessing whether the fleet is up. I turned down a second competing tool that a peer group used. It met none of the real constraints and would only have added one more thing to keep running.

Verify instead of trusting your gut about which system is under strain, because a label and reality can point in opposite directions. I reorganized the system around what it does, not around habit or an assumed label.

Story details

All stories · Work record