Day 11: What happens when your AI agent hits its usage limit? Midjourney drew 211 frames and nobody checked them

Day 11 is the day my crew went home and the photocopier kept running.

Day 11 at a glance

What happens when an AI agent hits its usage limit in the middle of a production? On my film, on October 14, everything that needed Claude to think stopped: the edit room, the posting schedule and the frame checks. The Midjourney runner kept going, because it's a script that only submits prompts and saves the results. By midnight it had drawn 211 frames for Scenes 11, 12 and 13. Nobody looked at any of them that day. RunPod spend for October 14: $0.

How did the crew run out?

Claude works on a plan with a weekly limit. October 13 was a heavy day: the frame fixes, the sound pass, sound for 51 future scenes, the social catch-up and two fact checks. Late that night the scheduled jobs started failing with the same message: weekly limit reached, resets October 16 at noon.

The scheduled jobs don't know that. Every two hours the edit room woke up to improve a scene and failed within seconds. On October 14 that happened 12 times in a row. The diary posting job failed at all three slots, so the posts for that day never went out, and the diary fell a day behind.

Why did Midjourney keep going?

Since Midjourney flagged the account on Day 02, frames come from a paced runner. It's a short Python script. It takes the next shot in the queue, submits the prompt with the character references, waits, saves four takes, and moves on. It doesn't need Claude for any of that.

So on October 14 it did exactly what it was built to do. After a few jobs past midnight it slept until 8:55 in the morning, never ran more than 4 jobs an hour, took breaks of 15 to 44 minutes, and finished its last job at 11:51 at night. That added up to 53 jobs, 4 takes each: 9 shots from Scene 11, 19 from Scene 12 and 25 from Scene 13.

The pacing rules worked. Midjourney didn't complain once. What the script can't do is look.

What did nobody catch?

Around noon on October 15, Claude had a short window and ran its pre-judge on the backlog. Every new frame gets checked against the character locks, the script and the style before I see it. 108 of the 211 failed.

65 of those failures were the same note: John's hair. John's lock has light sandy hair. Under our orange light, Midjourney kept giving him dark brown or near-black hair. We had spent October 13 fixing that same hair, frame by frame, in older shots.

Some failures were stranger. These are Claude's notes, from its check:

• In Scene 12, John "has grown a goatee/beard and near-black hair - looks like a second Mike." Three takes in a row.
• In Scene 13, the man attacking John is "drawn as a second John," and the real John lies on the floor as a huge head on a shapeless body.
• In another Scene 13 shot, John appears twice: a big John up front and a second one walking away behind him.
• In the uppercut shot there is "No uppercut: John stands passive with no fist raised while a tooth floats off the attacker cheek."

None of those reached me. That's the pre-judge's job, and it's why the runner's frames wait for it.

What did I approve?

I judged the 103 frames that passed on October 16 and approved 54. Most of the approvals are in Scene 12, John's flash of memories, where he stares out from in front of a wall of TV screens, and in the Scene 13 fight.

The one with a note is Scene 13, shot 22. John takes a plain white card from a dead man's pocket and holds it up to the light. The script says it's printed with the Megaplex mark, a circle crossed by three lines. Midjourney drew four takes. Claude's check failed three of them: one card had a single slash, like a no-entry sign, and two had an X. The fourth had an N inside the circle. Claude passed that one with a warning that I should check the symbol. I approved the shot and wrote:

"need to be 3 vertical lines in a circle not an N"

The pose and the light are right, so the card gets fixed later. Four takes, four different symbols, and none of them was three lines.

The numbers for October 14

• Claude: weekly limit reached late on October 13; back for a short window on October 15; full reset October 16 at noon.
• Edit room: 12 scheduled runs, 12 failures. Posting: 3 slots, 0 posts.
• Midjourney runner: 53 jobs, 211 frames (Scenes 11 to 13), no warnings.
• Claude's check, a day later: 103 passed, 108 failed, 65 of them for John's hair.
• My verdicts, two days later: 54 approved, 49 rejected.
• RunPod: $0.

So what did Day 11 teach me?

The parts of the pipeline that are plain scripts kept working, and the parts that need judgment stopped. That's the right way around. A runner that draws without anyone checking is only safe because nothing it makes goes anywhere until something with eyes looks at it.

It also showed me where the real bottleneck is. Midjourney can make 211 frames in a day. Half of them were wrong in the same way, and 65 of them had the same wrong hair. Fixing that in the prompt is cheaper than catching it 65 times. Follow the run on the Making Death Match series page.

FAQ

What happens when an AI agent hits its usage limit?
On Day 11 every scheduled job that needed Claude failed: 12 edit-room runs and 3 posting slots. Plain scripts that don't call the AI, like the Midjourney runner, kept working.

How many Midjourney frames can a paced runner make in a day?
On October 14, 53 jobs and 211 frames, with no jobs before 8:55 a.m., at most 4 jobs an hour and breaks between batches.

Why did so many AI frames fail the next day's check?
108 of 211 failed. 65 of those had John's hair drawn dark instead of his light sandy lock. Others doubled characters or broke the script.

Did the unchecked frames reach the director?
Only the 103 that passed Claude's check. I approved 54 of them on October 16.

While you wait for the next update: play Halfsies

Halfsies is our iPhone game, live now. One cut. Two halves. Each day you get three objects and one attempt at each: cut it exactly in half by eye and the scale tells you how close you got. It runs fine without an AI crew. Download Halfsies on the App Store