The local machine
52 plates, 18 sprites
Plates from 14:49 to 15:30 and sprites from 15:08 to 15:35, with 48 ambience beds between 15:32 and 15:35.
Real-time film · 29:42 · production notes · 8 October 2026
From gears to silicon
A twenty-nine-minute, forty-two-second Art Deco film about two thousand years of machines that calculate, from the bronze Antikythera mechanism to chips with billions of switches and machines that learn, with every machine drawn working. Thirty agents built it in under six hours, one animator for each chapter. This is where the time and the money went.
The session in four numbers
From a handoff brief to a published film
5 h 53 minfrom the brief at 14:11 to the final report at 20:04, with one 20-minute pause
32 h 25 minof agent working time, run in parallel
30 agentsone Opus director, 16 Sonnet and 13 Haiku agents
$224.56if every call had been billed at API list prices
29:42 film14 chapters68 shots128 lines, 4,090 words51 plates18 sprites48 ambience beds19,000 lines of film code10 research notes
The day
14:11 to 20:04
Each bar is one agent, from its first instruction to its last report; each thin mark is a tool call. The four local rows mark every picture, sprite, voice line and ambience bed the machine produced. Research went first, eight chapter animators started together, a second wave followed, and the fix agents came last.
Opus 5.5Sonnet 5.5Haiku 5.5Local machine, no API cost
The company
Who did what, for how long, at what price
| Agent and assignment | Ran (local) | Time | API calls | Tool calls | Cost at list prices |
|---|---|---|---|---|---|
| Director · Opus 5.5 | |||||
| Director: orchestration, review and export | 14:11–20:04 | 5 h 53 min | 595 | 651 | $88.86 |
| Research · Haiku 5.5 | |||||
| Research: inside a CPU, measures | 14:18–15:19 | 1 h 01 min | 23 | 70 | $0.58 |
| Research: learning, perceiving machines | 14:18–14:52 | 34 min | 15 | 79 | $0.46 |
| Research: counting, clocks, automata | 14:18–14:52 | 34 min | 13 | 60 | $0.58 |
| Research: relays, tubes, first computers | 14:18–15:06 | 48 min | 21 | 68 | $0.74 |
| Research: logic, binary, switching | 14:18–14:53 | 36 min | 23 | 65 | $0.46 |
| Research: mechanical calculators | 14:18–15:00 | 42 min | 14 | 64 | $0.76 |
| Research: silicon, transistor, chips | 14:18–14:55 | 37 min | 24 | 80 | $0.67 |
| Research: Antikythera mechanism | 14:18–15:12 | 54 min | 20 | 50 | $1.09 |
| Research: programmable machines | 14:18–15:08 | 50 min | 22 | 75 | $0.71 |
| Research follow-up: Mars, AI facts | 14:59–15:18 | 18 min | 9 | 42 | $0.20 |
| Source checks · Haiku 5.5 | |||||
| Verify Difference Engine No. 2 works | 15:13–15:25 | 12 min | 8 | 14 | $0.09 |
| Verify remaining small descriptors | 15:57–16:11 | 14 min | 12 | 30 | $0.11 |
| Compile the film's sources panel | 15:41–15:55 | 14 min | 10 | 26 | $1.07 |
| Animation, one agent per chapter · Sonnet 5.5 | |||||
| Animate chapter IV (holes in cards) | 16:35–18:07 | 1 h 32 min | 145 | 158 | $9.28 |
| Animate chapter VIII (solid switch) | 16:35–18:00 | 1 h 25 min | 171 | 187 | $9.96 |
| Animate chapter IX (grey crystal) | 15:24–16:32 | 1 h 09 min | 132 | 147 | $7.75 |
| Animate chapter V (true or false) | 15:24–16:35 | 1 h 11 min | 111 | 141 | $7.50 |
| Animate chapter XI (city on a chip) | 15:24–16:43 | 1 h 20 min | 155 | 176 | $9.72 |
| Animate chapter XIV (perceive, act) | 16:35–18:09 | 1 h 34 min | 140 | 195 | $10.78 |
| Animate chapter VI (one machine) | 16:35–17:42 | 1 h 06 min | 125 | 138 | $8.34 |
| Animate chapter XIII (learning) | 16:35–17:56 | 1 h 21 min | 156 | 163 | $9.11 |
| Animate chapter II (the tick) | 16:18–17:25 | 1 h 06 min | 126 | 138 | $6.71 |
| Animate chapter III (the carry) | 15:24–16:48 | 1 h 24 min | 155 | 166 | $10.52 |
| Animate chapter VII (glass and fire) | 15:24–16:18 | 54 min | 141 | 167 | $6.53 |
| Animate chapter X (printed with light) | 15:24–16:31 | 1 h 07 min | 137 | 156 | $6.95 |
| Animate chapter XII (measures) | 15:24–16:30 | 1 h 06 min | 150 | 161 | $7.79 |
| Animate chapter I (Antikythera) | 15:24–16:28 | 1 h 05 min | 133 | 140 | $8.02 |
| Fixes · Sonnet 5.5 | |||||
| Fix chapter II verge and tag | 17:26–18:21 | 55 min | 119 | 126 | $5.72 |
| Make Difference Engine 2 compute | 18:09–18:52 | 42 min | 84 | 90 | $3.50 |
| 30 agents | 2,989 | 3,823 | $224.56 | ||
The bill
Where $224.56 went
By role
Director · $88.86 · 40%Engine, script, score, plate fixes, review of every chapter, export and publishing.
10 Haiku researchers · $6.25 · 3%One research note each, from the Antikythera mechanism to the AI follow-up.
3 Haiku checkers · $1.27 · 0.6%Descriptor checks, the Difference Engine check and the sources panel.
14 chapter animators · $118.96 · 53%One Sonnet agent for each of the fourteen chapters.
2 fix agents · $9.22 · 4%Chapter II's verge and tag, and a working Difference Engine No. 2.
By model
By kind of token
The director, hour by hour
Each gap between two transcript events is given to the tool call or message that ended it, then sorted by tool name, file path and command. The buckets are keyword matches, so read them as proportions. 5 h 06 min in all.
The animators
How long each chapter took
ch XIV (perceive, act), the longest, ran 1 h 34 min. Each chapter had its own contact sheet for the director to review.
Fourteen chapters, one animator each
The film
What the day produced
Narration
128 lines
4,090 words, almost all read by af_bella, with three lines for guest voices. The voice lines were produced between 15:16 and 16:12.
The score
One ticking pulse
A continuous score on a ticking pulse and a turning arpeggio, growing from lyre and music box to harpsichord, relays, valves, synthesiser sequences and full orchestra. The first mix measured speech at −21.8 dB with a peak of −1.0 dB.
The MP4
29:42, 1.1 GB
1376 by 768 at 30 fps, exported in 20 chunks, with an 83 MB single-file page. A scan found no script errors and no missing letters, and frames at 8, 22 and 28 minutes matched fresh renders.
What the checks caught
Eight problems found in review
Every chapter was reviewed from contact sheets before it was accepted. These are the failures the review recorded, with how each was fixed.
| Caught in review | What went wrong | How it was fixed |
|---|---|---|
| Plates: Huygens, Lovelace, gpu_hall, mars_dusk, Turing | An anachronism and wrong numerals on Huygens, 1910s dress on Lovelace, a statue in the GPU hall, caped figures on Mars, identical poses and green skin on Turing | Reprompted with new seeds until each was right |
| Fact audit | About 35 wordings that claimed too much or too little, found by a sources agent and a descriptor check | Script lines corrected, for example Su Song’s clock “more than nine metres”, the Z3 “destroyed in an air raid” in 1943, the ENIAC programmers named, and “a chip is not simply made of sand” |
| Perceptron and language models | The first public demonstration and the way chatbots answer were stated loosely | “First shown in 1960” and “tuned for conversation, they answer in our own words” |
| Engine: group transform | GEARS.group ignored the full current transform, which broke the opening and finale | Fixed in the engine and noted in ENGINE.md |
| Score level | High crest factor that would have triggered the mixer’s normalisation, and uneven chapter levels | A soft limiter and a gain per chapter |
| Narration | Several lines read too fast | Speed set to 0.9 or 0.85 on those lines |
| On-screen layout | Clipped title, clipped sky ring, a rover glint overlapping a plaque in the finale | Each fixed and re-checked on frames |
| Dates and labels | Several on-screen dates and labels had drifted from the narration | Checked against the script line by line; the lesson went into the handover notes |
What got in the way
The tools fought back
Four things outside the film slowed it down or put other work at risk.
| What got in the way | What happened | What changed |
|---|---|---|
| The export tool | On finishing, it closed every export browser on the machine. When Cathedrals of Light’s MP4 finished, it killed this film’s export | It now closes only its own browser, and the film was exported in 20 short chunks that can each be redone |
| The shared image queue | Another session’s worktree occupied ComfyUI, so a second build was refused | A --share-queue option, so builds wait their turn |
| Disk space | C: had 45 GB free that afternoon | Scratch files, temporary folders and the preview clip were deleted, leaving 312 GB free |
| Shell quoting | Heredoc quoting failures when patching files | Patch scripts written with the file tool and then run |
What it taught
Five things the numbers say
One animator per chapter
Fourteen chapters had fourteen Sonnet animators at $118.96, an average of $8.50 each. Wave one of eight started at 15:24, chapter II at 16:18 and a second wave of five at 16:35.
Research ran wide and finished early
Ten Haiku researchers cost $6.25 between them and the last finished at 15:19, before the first animator started.
The director waited for agents longer than it did anything else
1 h 39 min of the director’s hours went on waiting for agents, more than the export (43 min) or the score (34 min). The last chapter, ch XIV (perceive, act), reported at 18:09.
Checking found the errors that mattered
The fact audit changed about 35 lines and the review caught plates that were wrong in ways a contact sheet shows at once. The lesson recorded afterwards was to check every on-screen date and label against the narration.
The cheapest film per minute so far
At $7.56 per finished minute it undercuts the others measured here ($13.05 for Promethean Redrawn, $13.18 for Cathedrals of Light, $50.83 for the original Promethean). The brief had said not to worry about cost, and the structure kept it down anyway.
Method
How these figures were measured
Measured from the session's own transcripts: the director's, and one for each of the 29 agents it launched. Assistant messages are deduplicated by message id and the last entry's usage is kept. Models come from each agent's metadata. The session was summarised once, at 16:31, and the director’s figures include both halves.
Cost is what the usage would have cost at API list prices in October 2026: Opus 5.5 at $4 per million input tokens and $20 per million output, Sonnet 5.5 at $2 and $10, Haiku 5.5 at $0.10 and $0.50 (higher rates above a 100,000-token prompt), cache reads at 0.05 times input, cache writes at 1.25 times (five minutes) or 2 times (one hour). A subscription session was not billed this way, so treat the dollar figures as a measure of work, not an invoice.
Times are UK local time (BST, UTC+1). Local-machine marks are the modification times of the generated files. The director’s transcript has a 20-minute pause between 19:39 and 19:59. The script text, the audit and the export problem come from the session’s own summary and final report. The hour-by-hour split sorts tool calls by keyword, and gaps of more than 20 minutes with no events are left out of it.