The Allowance Problem, Day 3: More Supervision Doesn’t Equal More Judgment
Anthropic gave its AI shopkeeper a boss.
Read field note →Field Notes archive
All writing
Long-form notes on the systems, decisions, and controls that make technology useful.
The Allowance Problem, Day 2: The AI Boss Didn’t Save It
An AI agent ran a real shop, with an AI boss watching over it. It still ended the month ordering a PlayStation 5 and a live fish.
Read field note →The Allowance Problem: Credit Card or Debit Card?
We already ran the experiment: should AI agents get a credit card or a debit card?
Read field note →The Allowance Problem
Would you give your AI agent a credit card or a debit card?
Read field note →The operating model for the capability era
the models are improving faster than most enterprises are improving their control systems. The real gap is operational control, and the answer is better systems around the models.
Read field note →AI spend is becoming an operational risk category
AI cost is no longer just a procurement conversation. If agent usage grows faster than your visibility into spend and business value, you do not have control. You have drift.
Read field note →The attack surface is now the trust surface
if you can poison the logs an agent reads, you can influence what it does. The attack surface is now the trust surface, which makes data trust, observability, and identity security the same conversation.
Read field note →Autonomous by default is a governance decision
when autonomous mode becomes the default, that is not just a product feature. It is a governance decision about identity, permissions, approvals, and observability.
Read field note →Better model. Bigger control problem.
the models are getting better, but that is no longer the interesting problem. Capability is scaling faster than control, and that gap is where enterprise teams get hurt.
Read field note →