The Beta Caught the Wrong Question

The fifth 1.7 beta had passed its artifact checks. The next change looked small: give every tool call a human-readable line so a person could follow an agent without opening raw arguments.

Our first contract named that line reason. It worked mechanically. The model filled the field, Core transported it, and the React client could read it. Then we put the result in the conversation UI.

A simple system check no longer looked simple. Above uname -a, the agent wrote that it needed to inspect the operating system because the user had asked about release readiness. The interface repeated the prompt, claimed to know the agent's motivation, and used more space explaining the call than reporting its result. The field had answered exactly the question we asked. We had asked the wrong question.

That was the moment the beta earned its name. Compatibility tests could prove that a string survived the wire, but only the end-to-end reader showed that the string did not belong in the product.

For Beta 6, the field is summary: a short action phrase such as “Check the operating system.” The public React beta learned to preserve it first. O Chat then moved the action beside a status dot and quiet duration, with the tool name, arguments, and result folded underneath. Only after that reader reached production did Core begin writing summaries into live and replayed traces.

Older calls still execute without the field. Ordinary functions never receive the presentation value as a surprise argument, while a function that genuinely declares its own summary parameter keeps it. Those details matter, but they are not the lesson of the beta.

The lesson is that a contract can be type-correct and still ask the product to lie. Beta 6 changes the question from “Why did you do this?” to the one the interface can answer honestly: “What are you doing?”