The local machine
Generated and edited from 22:38 to 00:32 (98 plates and 48 cut-out figures), each derived into toned paper layers with ink-order and wash-order reveal maps. Plus 42 Stable Audio effects, finished between 23:52 and 00:27.
A remake of Promethean that keeps its toned paper, its ink and its single flame and grows the film from 2:51 to 22 minutes 42 seconds, with drawn plates and figures behind the models, a narrator and a score of nine movements. Thirty-three agents built it in a little under three and a half hours. This is where the time and the money went.
Each bar is one agent, from its first instruction to its last report; each thin mark is a tool call. The three local rows mark every picture, sound and voice line the machine produced. The research ran first, then the shared engine, then 19 animators at once, then the director’s own passes.
| Agent and assignment | Ran (local) | Time | API calls | Tool calls | Cost at list prices |
|---|---|---|---|---|---|
| Director · Opus 5.5 | |||||
| Director: orchestration, review and export | 21:51–01:24 | 3 h 33 min | 495 | 600 | $63.70 |
| Engine, typography and plate pipeline | |||||
| Study Promethean engine code | 21:52–22:01 | 9 min | 38 | 54 | $1.50 |
| Build typography module and fonts | 22:15–22:51 | 36 min | 142 | 168 | $5.99 |
| Build plate derivation pipeline | 22:16–23:01 | 45 min | 113 | 154 | $5.16 |
| Engine core · Opus 5.5 | |||||
| Build Promethean engine core | 22:15–23:03 | 48 min | 162 | 197 | $17.60 |
| Score · Opus 5.5 | |||||
| Compose the Promethean score | 22:39–00:40 | 2 h 01 min | 121 | 173 | $18.47 |
| Research · Haiku 5.5 | |||||
| Research 01 myth and cosmos | 22:10–22:17 | 7 min | 12 | 50 | $0.16 |
| Research 02 Greece | 22:10–22:22 | 12 min | 20 | 53 | $0.21 |
| Research 03 Rome | 22:10–22:21 | 11 min | 18 | 52 | $0.23 |
| Research 04 keepers and cathedrals | 22:10–22:22 | 12 min | 17 | 58 | $0.24 |
| Research 05 Renaissance | 22:10–22:22 | 12 min | 30 | 65 | $0.24 |
| Research 06 Enlightenment | 22:10–22:20 | 10 min | 18 | 55 | $0.24 |
| Research 07 industry | 22:10–22:22 | 12 min | 17 | 59 | $0.26 |
| Research 08 modern and data | 22:10–22:22 | 12 min | 23 | 58 | $0.19 |
| Animation, prologue and coda · Opus 5.5 | |||||
| Animate a00 prologue | 23:06–00:25 | 1 h 20 min | 186 | 238 | $19.53 |
| Animate a18 coda | 23:07–00:19 | 1 h 12 min | 182 | 196 | $16.06 |
| Animation, seventeen parts · Sonnet 5.5 | |||||
| Animate a01 Greece part one | 23:06–00:25 | 1 h 19 min | 115 | 126 | $6.95 |
| Animate a02 Greece part two | 23:06–00:02 | 56 min | 115 | 144 | $6.00 |
| Animate a03 Rome roads and water | 23:06–00:13 | 1 h 07 min | 110 | 137 | $8.17 |
| Animate a04 Pantheon and the fall | 23:06–00:22 | 1 h 16 min | 145 | 158 | $10.76 |
| Animate a05 Iona | 23:06–00:17 | 1 h 12 min | 161 | 228 | $8.85 |
| Animate a06 relay and cathedrals | 23:06–00:04 | 58 min | 114 | 132 | $7.09 |
| Animate a07 Florence and perspective | 23:06–00:40 | 1 h 34 min | 131 | 166 | $10.56 |
| Animate a08 press and globe | 23:06–00:24 | 1 h 18 min | 129 | 144 | $8.97 |
| Animate a09 Copernicus and the Globe | 23:06–00:31 | 1 h 25 min | 132 | 185 | $9.26 |
| Animate a10 Newton and music | 23:07–00:28 | 1 h 22 min | 146 | 170 | $8.45 |
| Animate a11 1776 and Jenner | 23:07–00:30 | 1 h 23 min | 217 | 265 | $11.44 |
| Animate a12 Watt and machines | 23:07–00:19 | 1 h 12 min | 162 | 188 | $10.79 |
| Animate a13 Maxwell, light, voices | 23:07–00:23 | 1 h 16 min | 154 | 173 | $9.23 |
| Animate a14 Wright and Fleming | 23:07–00:29 | 1 h 22 min | 114 | 131 | $8.23 |
| Animate a15 skyscraper and atom | 23:07–00:14 | 1 h 07 min | 124 | 140 | $7.71 |
| Animate a16 the Moon | 23:07–00:22 | 1 h 15 min | 119 | 139 | $6.18 |
| Animate a17 silicon, network, pocket | 23:07–00:18 | 1 h 11 min | 139 | 160 | $7.68 |
| 33 agents | 3,921 | 5,016 | $296.11 | ||
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. 3 h 33 min in all.
Every part of the film had its own agent and its own contact sheet of 12 to 14 stills for the director to review. The shortest ran an hour or less; the longest set the clock for everything that needed all of them.
Generated and edited from 22:38 to 00:32 (98 plates and 48 cut-out figures), each derived into toned paper layers with ink-order and wash-order reveal maps. Plus 42 Stable Audio effects, finished between 23:52 and 00:27.
2,841 words read by bm_fable at speed 0.88. The voice lines took about two minutes of CPU time at 22:35, and every shot sits on a bar line of its own movement’s tempo.
The original’s synthesiser grown into 640 cues. Final mix: bed −20.8 LUFS with a peak of −1.0 dB, speech −21.4 dB, no clipping.
1.04 GB, exported in chunks on the GPU in about 20 minutes. Chosen at 60 frames a second because fine lines judder at 30. A 462 MB and a 99 MB copy were made for sharing.
Each of the 19 parts was reviewed from stills before it was accepted, and at least twelve of them went back for a polish round. These are the failures the review recorded, with how each was fixed.
| Caught in review | What went wrong | How it was fixed |
|---|---|---|
| Appia milestone | Garbled pseudo-inscription generated into the stone | Edited out; face left as plain weathered stone |
| Florence | The dome came out finished, but the shot builds it | Re-edited to an open drum with scaffolding and a crane |
| Moon horizon | Came back photoreal | Regenerated as a chalk drawing |
| Padua roof | A telescope and modern aerials on a 1610 roof | Removed with an edit of the edit |
| Workshop and bench plates | A painted press and an engine already in the picture | Removed, so the models can arrive on the plinth |
| Bertha and Wilbur figures | A car wheel in one, a faint wing in the other | Both re-edited without them |
| Galileo figure | A cream triangle between the tripod legs, from filling every hole in the mask | Fill limited to holes under 600 pixels |
| Cards over busy plates | Words illegible against the drawing | A stronger halo in the type module, for every card |
| Night plates | Bare cards pale before their wash | Each card carries its own paper colour |
| Engine | Text parts dropped their tag; static parts ignored shadow:false | Both fixed in the engine |
| Promethean | Promethean Redrawn | |
|---|---|---|
| Film | 2:51, 1080p at 30 | 22:42, 1080p at 60 |
| Session | 2 h 08 min | 3 h 33 min |
| Agents | 7 (all Opus) | 33 (5 Opus, 20 Sonnet, 8 Haiku) |
| Agent working time | 8 h 47 min | 33 h 07 min |
| Cost at list prices | $145.04 | $296.11 |
| Cost per finished minute | $50.83 | $13.05 |
| Session minutes per finished minute | 44.7 | 9.4 |
| Generated media | none | 146 pictures, 84 voice lines, 42 effects |
The 19 animators cost $181.92 of $296.11, or 61%: $146.33 for 17 Sonnet agents and $35.59 for the two Opus ones. Each ran inside a context of 400,000 to 620,000 tokens, and about 62% of the whole bill was cache reads.
The engine core, the typography module and the plate pipeline finished at 23:03, 22:51 and 23:01. The 19 animators started at 23:06, each with an engine that already drew plates, figures and type, so none of them had to solve the same problem twice.
The animators ran from 56 min to 1 h 34 min. The director spent 1 h 19 min waiting on agents, the largest single block of its time, and could not start the final score render, the continuity sheet or the export until the last part, Florence and perspective, reported at 00:40.
Eight research notes cost $1.77 together, about 0.6% of the total, and every one finished in 7 to 13 minutes, before the animators needed them.
Asked whether 60 fps would cost extra coding, the director answered that it would not: the film is a pure function of time, so the exporter reads the rate from the storyboard. The real cost was in the export, with twice the frames and a bigger file.
The director’s context filled once, at 00:50, and the session was summarised. DIRECTOR_TODO.md, kept current for this purpose, carried the open items across, and the director went on to retime cards, re-render the score with all 640 cues and start the export.
Measured from the session's own transcripts: the director's, and one for each of the 32 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.
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). The session began at 21:51 on 10 October and its last transcript event is at 01:24 on 11 October. Local-machine marks are the modification times of the generated files. Export duration and file sizes come from the project’s own notes and files. The hour-by-hour split sorts tool calls by keyword.