I went to bed last night with a list of things that needed doing. When I woke up, six of them were done.

This is the strangest part of making a movie with AI agents. You set up tasks, go to sleep, and wake up to pull requests. Not vague progress reports. Actual code changes, merged and deployed, with commit messages and everything.

What Happened Overnight

The assistant director—a Claude instance running on the task server—coordinated a batch of remote agents through the night. By the time I had coffee, six PRs had been reviewed and merged:

  • The Feb 7 character lock blog post went live on the site
  • The characters page got updated with all seven approved turnaround sheets
  • Act 1 storyboards were switched from broken local paths to R2 URLs, so they actually display now
  • Hero images generated by Gemini were added to the homepage
  • Scene 1 panels were regenerated using image-to-image with the approved character refs
  • The storyboard viewer got fixed and the index page updated

That last one matters more than it sounds. The storyboard viewer had been broken since we migrated images to R2 storage. Navigation was busted, image paths were wrong, and half the panels showed broken image icons. Now it works. You can actually click through Act 1 and see the story.

27 Storyboard Panels Regenerated

The bigger overnight effort was the storyboard consistency pass. Remember yesterday's post about character lock? The whole point of locking designs was to use them as references when regenerating panels where characters had drifted off-model.

The agents ran a consistency audit across Scene 1 and regenerated 27 panels using the approved turnaround sheets as image-to-image references. Each panel was checked against the locked designs, flagged if characters didn't match, and regenerated with the correct references fed into Gemini.

The results are noticeably better. Gabe actually looks like Gabe in every panel now, not like four different middle-aged men wearing the same shirt. Nina's proportions stay consistent. Leo keeps his dinosaur pajamas.

Is it perfect? No. AI-generated storyboards still have variation between panels. But we went from "is that even the same character?" to "that's clearly the same person in a different pose." For rough storyboards, that's good enough.

First 3D Character Models

The overnight session also kicked off the first attempt at converting our 2D character designs into 3D models. Gabe and Nina went through the TRELLIS pipeline—feeding their approved turnaround sheets into AI 3D generation and getting back mesh files.

The results are... preliminary. They have the right general shape and proportions, but the topology is rough and the textures need work. Think of them as 3D sketches, not production-ready models. We'll iterate from here.

But having any 3D representation of these characters is a milestone. It means we can start thinking about scene layout, camera angles, and lighting with actual geometry instead of just flat storyboard panels.

What Worked

The autonomous task pipeline worked better than I expected. Tasks got created, agents picked them up, PRs got opened, code got reviewed (by the assistant director), and merges happened. The whole loop ran without me touching anything.

The key was having clear, small tasks. "Fix the storyboard viewer" is a good task. "Make the website better" is not. Each PR was focused on one thing, which made automated review feasible.

What Didn't

Merge conflicts. When multiple agents work on the same files simultaneously, you get conflicts. The assistant director had to resolve several of them, and its merge resolution isn't always clean. A couple of times it chose the wrong side of a conflict and had to be corrected.

Also, the agents don't have great taste. They can regenerate a storyboard panel to match a reference, but they can't tell you whether the composition serves the story. Technical correctness and creative quality are different things. The consistency pass improved character likeness but didn't improve storytelling. That still needs human eyes.

What's Next

The overnight session proved the workflow. Now we scale it up. The consistency audit needs to extend beyond Scene 1 to cover all of Act 1—that's scenes 2 through 10. The 3D models need iteration. And we need to start thinking about social media—we've been so heads-down on production that we haven't told anyone this project exists.

But tonight, I'll go to bed knowing the agents will keep working. That's still weird. In a good way.