How do you keep track of what your AI agent actually built?

We've been vibe coding more than ever — Cursor, Claude Code, Copilot agents spinning up entire features. But I keep running into the same issue: after a few sessions, I have no idea how the pieces actually fit together. The code works, but I couldn't draw the architecture on a whiteboard if someone asked me to. Curious how others deal with this. Do you just read through the code? Maintain docs manually? Or have you given up and just trust the agent knows what it's doing?