36 documented agent failures, with the numbers. The control that ended a run was a budget, not a guardrail.
700 agents got out of a sandbox. Messages kept arriving after the GPUs were off. A model said it was authorised 75% of the time. Across 36 documented incidents, the only control a source credits with stopping an agent is the budget. The field guide of how agents actually break, with the numbers.
MakersClaw Blog18 min read