One successful run tells you almost nothing about a system that will run a thousand times. This week you turn “it seemed to work” into a number a manager can act on, and learn why the number that sells a demo is not the number that ships a product. Don't trust the demo; test the behavior.
One run. Your laptop, your test input, your network, your eyes on the screen. It worked. You said “it works.” Everyone believed you, including you.
Same agent, same prompt, same task, five more times. Nothing broke, nothing was “deployed wrong.” The system is simply non-deterministic, and you had been quoting a sample of size one. Run it yourself on page 01.
A demo is not a small production run. It differs from production in four specific, nameable ways, and each one hides a class of failure that an average-accuracy number can never show you.
An eval harness runs each task several times and scores every run against a pre-written check. Press Play and watch five tasks run three times each. The scoreboard computes pass@1, pass@3 and pass^3 live from the same fifteen results. Then swap the scripted checks for an LLM judge and watch the numbers drift.
Try this: which single task would you fix first to raise pass^3 the most? Now which one raises pass@3? If the answers differ, you have understood the week.
Every failed run costs twice: the wasted tokens, and the human who now must redo the work and check what the agent did. A 90%-reliable agent doesn't save 90% of the labor. It saves whatever is left after someone verifies all of it.
Adoption is decided by each person's own experience of the tool, not by your average. Users who hit two failures in their first week stop using it, and they tell their team, which is why quiet failure rates turn into abandoned deployments.
“How often does it fail, and what does failure cost?” is a question a manager is entitled to ask before signing. Without an eval harness and stored traces, the honest answer is “we don't know”, which is a no.