Read the Tool Call, Not Just the Answer

You ask Cowork a question. The reply looks good. You move on.

A week later, you find out the answer was based on the wrong source: Cowork pulled from a deprecated doc instead of the current one. Or it answered your "what did I commit to last week" using your sent folder but never checked the threads where the actual commitments were made. The answer wasn't wrong because the model was bad. It was wrong because Cowork looked in the wrong place, and you didn't notice.

The fix is small and immediate: read the tool calls.

#What "tool calls" means here

When Cowork answers a question that involves your connectors or files, it doesn't make the answer up. It runs tools (fetches your calendar, queries your email, reads a doc, searches a folder) and uses what those tools return to write the reply.

These tool runs are visible. They appear in the response as small entries showing which tool was called and with what input. They're easy to skip over if you're focused on the final answer; they're also where 80% of trust calibration lives.