If you’ve built AI agents for clients, how do you prove the agent actually did the work?

I’ve been thinking about this while working on an agent-to-agent workflow. Most agent de mos seem to end when the agent produces an answer or completes a task. But for real client work, I’m not sure that’s enough. If Agent A passes work to Agent B, or several agents work together, how do you show the client what actually happened? Do you just give them logs and screenshots? Do they ever ask which agent version ran, what tools or files it used, what permissions it had, or whether a human checked the final result? I’m especially curious whether missing proof has ever delayed a project being approved, deployed or paid for. I don’t have a product link to share. I’m trying to find out whether this is a real problem outside my own head. If you’ve shipped an agent or MCP workflow for a real client, how did you handle this?