Every Mediovsky piece tagged AI agents.
Roundups frame agentic commerce as a standards war between OpenAI and Google. Read both specifications and OpenAI’s retreat from Instant Checkout, and the two rails are converging on one journey, asking merchants for the same feed at a cadence the spec never demanded.
Every team now has models writing, classifying, and summarizing at volume, and almost none can say whether the output got better or worse last month. The fix is an eval: a frozen set of real inputs, a rubric that discriminates, and a threshold. The research on how to build one is unusually clear, and unusually uncomfortable about the shortcut everyone takes.
When a single AI agent hits its ceiling, the work moves to a fleet with an orchestrator on top. The 2026 problem is no longer model quality. It is coordination, guardrails, and reliability.
