Control before, proof after: an accountability primitive for AI agents

There's a pattern I kept seeing. A team gives an agent real capability, like moving money, shipping a change, or resolving a ticket that touches a customer's account. For a while it's great. Then the ...