An Action Is Not a Reason

We asked an agent whether a machine was ready for release. It called inspect_system with uname -a, and the first version of the interface put a field named reason above the call.

That small name changed what the model wrote. Instead of a crisp activity such as “Check the operating system,” the transcript began explaining itself: “I need to inspect the operating system because the user asked me to verify release readiness.” The sentence was longer than the useful result, repeated the user's request, and presented a generated account of motivation as though it were an observable fact. On a phone, it also turned one quiet tool event into a block of prose.

The failure was not in the tool. The function still received the right command and returned the right kernel version. The failure was in the contract we had given the model: ask for a reason, and it will try to justify itself.

The useful line was already hiding inside that explanation. “Check the operating system” says exactly what a person watching the run needs to know. So the 1.7 contract now calls the field summary and describes it as a short action phrase. The reader places that phrase beside the status and duration, then keeps the tool name, arguments, and raw result behind an expandable detail row.

Changing the name also forced us to decide what happens outside the happy path. Old recordings and third-party clients do not have a summary, so they still run and receive a deterministic label derived from the tool name. Ordinary tools do not receive presentation metadata as an unexpected argument. A tool that already declares a real summary parameter keeps its own value, while the trace still records the bounded activity phrase shown to the reader.

When we replayed the same release check, the transcript finally read like an activity log: “Check the operating system · 1.2s.” The command and its output were still one click away, but the default view no longer pretended to know the agent's inner motivation.

That is the distinction worth keeping: report what is happening, reveal the evidence on demand, and leave “why” to the conversation where it can be examined rather than asserted.