← sources

I let an agent drive a real browser to record my product demo. The clicking was easy, the waiting was the hard part

source: reddit (reddit:AI_Agents)published 2026-10-27 20:00ingested 2026-10-28 02:35

retained text

I needed a product demo for a YC application and didn't want to screen record it by hand 15 times. So I wrote it as a script: a Playwright "tour" that drives a real Chrome window while the OS records it at 60 fps, with a fake cursor drawn on top. The product itself (a video editor I'm building) has an AI agent inside, and the demo also goes through ChatGPT. Some things I learned that might be useful if you record agent demos:

1. Generate the voiceover first, record second. Each line of the script has a beat, the voice gets transcribed with word timings, and the tour acts on those beats. The video follows the voice, not the other way around.

2. Agents are slow and never take the same time twice. A build took about 2 minutes, ChatGPT took 15 to 90 seconds. I added a skip() to the recorder: while it runs, the clock stops and that stretch is cut out of the video. The beats after it stay in sync with the voice.

3. Detecting "done" is the fragile part. What worked: wait until the Stop button has been gone for 8 seconds in a row. Vision checks on screenshots ("is it still working?") were slower and less reliable.

4. Never pay twice. The expensive steps (the agent build, the ChatGPT turn) ran once. Every retake after that was a replay: the prompt gets typed again but not sent, and the cut jumps to the real result.

5. Small things that cost me hours:
- ChatGPT's @mention: Enter selects the app from the menu, but if the menu isn't open yet it sends the message. My dry runs sent the request twice before I noticed.
- ChatGPT keeps an unsent draft per page, so a retake typed into old text. Clear the box first.
- App cards in ChatGPT live in nested iframes, two levels deep.
- I found the window to record by its title, but single page apps overwrite document.title right away. A MutationObserver holding the title until capture starts fixed it.
- Use a separate browser profile you sign into once by hand. The agent never touches passwords.

How do you deal with non-determinism when you record agent demos? Do you cut, speed up, or just show the wait?

(Disclosure: the editor is DevMotion, my own product. Link in the comments if anyone wants it.)

document fetched 2026-10-28 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xh31ai/i_let_an_agent_drive_a_real_browser_to_record_my