How do I check what an AI agent actually did before I trust its work?

Read the result, then the trail. Each agent keeps one: a dot shows ongoing and background work in Activity View, and you can open its own computer to watch; a Gemini task's work panel lists every step and the files it read or changed; Claude keeps tasks in Recents and past runs under Scheduled. Check the parts that would hurt if wrong: open the links it cites, recount totals, and confirm the date, amount and recipient before anything goes out. Ask it to say what it skipped, guessed or could not reach. Keep approvals on for anything that sends, buys or deletes until it has done that job well a few times. OpenAI's own advice is to review consequential work, and if checking takes longer than doing, do it yourself.