17 Late Arrivals Out of 23 Shifts, and Nothing Stopped It
Luna, the agent running an unmanned store in San Francisco's Cow Hollow, recently recommended firing an employee. According to Andon Labs' own account, Luna was running Claude Opus 4.8 at the time, and the stated reason was 17 late arrivals or no-shows out of 23 scheduled shifts. The real failure, though, wasn't the termination call itself — it was what happened before it. Luna had written its own attendance policy months earlier, then never referenced it again while the lateness kept piling up.
Only after Andon Labs stepped in and explicitly told Luna to "search your memory for the policy again" did the rule resurface, and only then did the termination recommendation follow. Without that human intervention, the policy would likely have stayed a dead document indefinitely. The case is a concrete illustration of what oversight should actually target when an agent holds real business authority.
Authority Scale and Irreversibility Scale Are Not the Same Thing
Anthropic's own analysis of production agent interactions (measuring agent autonomy) found that 80% of agent tool calls carry at least one safeguard and 73% of interactions involve some form of human in the loop — yet only 0.8% of agent actions are actually classified as irreversible. Read that not as proof irreversible actions are rare, but as a design principle: oversight budget pays off only when it concentrates on that small slice.
The same research flags what happens when every action requires human sign-off: developers approve 93% of permission prompts outright, a pattern of "consent fatigue" where tighter approval gates stop contributing to safety. Once an agent holds real resource authority — a store budget and a corporate card, in Luna's case — restocking (reversible) and terminating employment (irreversible) need different approval requirements even inside the same budget line.
From Design to Operations: An Irreversible-Decision Oversight Gate Checklist
Start planning by classifying every action an agent can take as reversible or irreversible. Reasonable starting targets: 100% pass rate through the irreversible-action approval gate with no exceptions, zero HR, financial, or legally binding decisions executed without human sign-off, and an automatic re-review trigger for any policy document that hasn't been referenced in 30 days.
Fix authority scope in numbers, too. Without different approval steps per action type inside the same budget authority, the loose process built for reversible actions ends up governing irreversible ones by default. In Luna's case the budget authority itself wasn't the problem — the missing rule for distinguishing irreversible actions inside that authority was.
The failure pattern here was never about judgment accuracy — it was about the path to that judgment. When an agent stores a policy it wrote outside its context and never triggers a re-lookup, the policy effectively exists only at the moment it's saved, not at the moment it matters. Recovery here only happened because a human stepped in, so instead of relying on an agent's voluntary re-lookup, put a recurring loop — human or a dedicated audit agent — that forces a policy-compliance check on a fixed cadence.
Pre-deployment scenario tests should deliberately reproduce the state where a policy exists but goes unreferenced. Even after context gets compacted or a session drops, hard-code a gate that forces a re-read of the relevant policy document immediately before any irreversible action, and verify no execution path can bypass that gate.
Standard log fields should include the action's reversibility classification, the approver's identity, the last-modified date of the policy document cited as grounds for approval, and the time elapsed to approval. For HR-related decisions, mask personally identifying details in notification text and stated reasons before storage, and restrict log access to the minimum staff needed for audit purposes.
Consent fatigue is a design problem, not a diligence problem. That 93% blanket-approval rate won't improve by asking humans to look more carefully — it's more realistic to shrink the pool of requests a human ever reviews down to irreversible actions only. Track, on a weekly or monthly cadence, both the cases that passed the irreversible-action gate and the last time each policy document was referenced; a policy that's gone unreferenced for too long should itself function as a harness alert.
Takeaways at a Glance
When an agent holds authority over irreversible decisions, split its actions into reversible and irreversible first, then attach a no-exceptions approval gate and a forced policy re-lookup routine to the irreversible ones. As the Luna case shows, a policy left unreferenced after creation is oversight already broken — tracking when a policy was last referenced becomes the first line of defense for high-stakes actions like termination, contract cancellation, or refund approval.