Watching a computer-use agent click through an interface on video makes automated operations look almost solved. The friction usually starts the moment you try handing that exact workflow over to run reliably inside a messy client setup. Most of the real challenge seems less about setting up the agent logic and more about keeping track of edge cases without sitting there watching it execute. For those running these for real businesses, where does the workflow usually break down first?

A few newbie questions this week asked where to run an agent: a VPS or a hosted sandbox. If you pick the VPS, these four steps prevent most of the breakage that shows up in threads here. 1. Run the agent as a service (systemd, or Docker with a restart policy), so a crash or a reboot brings it back without you. 2. Keep API keys in an env file outside the project folder, readable only by the user the agent runs as. 3. Turn on snapshots or backups before the first update. Updates are where most setups break, and a snapshot turns a bad one into a two minute rollback. 4. Pin the model id in the config, so a change of default on the provider side can't swap the model under you. What broke first on your setup?