An agent demo ends at the confirmation page; production starts at the first wall it is not allowed to pass. Typed stop reasons, an idempotent resume, an evidence log and a six-test matrix for the handover to a human at OTP, passkey, CAPTCHA, KYC and payment auth.
AI-written code often works for reasons nobody fully understands. How I pin its current behaviour with characterization tests in Laravel before refactoring: golden files, frozen ti...
An agent demo ends at the confirmation page; production starts at the first wall it is not allowed to pass. Typed stop reasons, an idempotent resume, an evidence log and a six-test...
Six lessons from building an MCP server that lets Claude, Cursor and Codex create and render images, PDFs and video, including the bug that hid a paid render from the agent that as...