Most of us here have shipped something that works beautifully while we sit and watch it run. This is about what changes the first night it runs on a schedule with nobody watching, because that is where the surprises actually live. Our guide to AI automation covers the whole picture, and these three are the ones that bite first.
The Model You Picked Is Now A Recurring Bill
With a human in the loop, model choice is a quality decision. Unattended, running a hundred times a night, it becomes a cost decision. An agent classifying 100 records makes 100 calls, and the premium model that felt obviously correct during development is now doing that on a timer while nobody reads the output closely enough to notice it was overkill.
The fix is boring. Match the model to the task rather than to the project. Classification and extraction usually do not need your reasoning model.
Guardrails Stop Being Paperwork
Cost ceilings, scope limits and output validation feel like compliance theatre when you are watching every run. Unattended, they are the only thing between a loop that misjudged its exit condition and a bill you find out about on Tuesday.
The two that earn their keep first are a hard spend cap per run and a scope restriction on what the agent is allowed to touch. Everything after that is refinement.
Failure Has To Handle Itself
The run that fails at 2:04am has nobody to retry it. That means tracking state between runs, so a failed batch is not silently skipped and not reprocessed twice on the next tick. It also means deciding what a partial failure halfway through a batch means, because the default answer is usually the wrong one.
Curious what broke first for everyone else here. Every person I ask has a different answer, and it is never the thing they planned for.
