The local machine
50 plates from 17:55 to 20:58 and 141 sprites, with 131 voice lines and 27 ambience and effects clips. Backcloths and cut-out pieces are generated; all motion, light, weather and type are drawn in code.
A nineteen-minute toy-theatre history of Scotland’s crown, told as the question of who has the right to wear it, with Mary, Queen of Scots as the longest act. An Opus director, nine Sonnet animators and ten Haiku researchers made it in an afternoon and evening. 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 row below marks every picture, voice line and sound the machine produced. Research ran first, then the animators in three waves, one act each, while the director built the score and mix. The gaps in the director’s bar are times with no activity.
| Agent and assignment | Ran (local) | Time | API calls | Tool calls | Cost at list prices |
|---|---|---|---|---|---|
| Director · Opus 5.5 | |||||
| Director: orchestration, review and export | 16:49–22:19 | 5 h 29 min | 647 | 702 | $91.94 |
| Research · Haiku 5.5 | |||||
| Research: Stewart kings to James V | 17:00–17:36 | 36 min | 32 | 107 | $0.52 |
| Research: traditional tunes in ABC | 17:06–17:53 | 47 min | 24 | 63 | $0.49 |
| Research: Bruce and independence | 17:00–17:53 | 53 min | 30 | 122 | $0.87 |
| Research: Mary, Queen of Scots | 17:00–17:48 | 48 min | 38 | 95 | $0.59 |
| Research: toy theatre and Stevenson | 17:06–17:48 | 42 min | 21 | 66 | $0.45 |
| Research: James VI and the Honours | 17:00–17:54 | 54 min | 25 | 114 | $1.24 |
| Research: Alexander III and succession | 17:00–17:43 | 43 min | 30 | 91 | $0.46 |
| Research: origins and Scone | 17:00–17:34 | 34 min | 13 | 59 | $0.66 |
| Research: Macbeth, Margaret, David I | 17:00–17:34 | 35 min | 19 | 79 | $0.58 |
| Fact check · Haiku 5.5 | |||||
| Fact-check five script claims | 18:02–18:29 | 27 min | 13 | 55 | $0.24 |
| Animation, one agent per act · Sonnet 5.5 | |||||
| Animate Act III part B | 19:01–20:32 | 1 h 30 min | 161 | 194 | $8.32 |
| Animate Act VI | 19:29–21:09 | 1 h 39 min | 188 | 204 | $13.02 |
| Animate Act IV part B | 19:01–20:18 | 1 h 17 min | 128 | 167 | $8.18 |
| Animate Act I: The Stone | 18:24–20:30 | 2 h 06 min | 221 | 247 | $17.56 |
| Animate Act V part A | 19:29–20:48 | 1 h 18 min | 146 | 171 | $9.71 |
| Animate Act III part A | 19:01–20:25 | 1 h 24 min | 151 | 169 | $10.54 |
| Animate Act IV part A | 19:01–20:50 | 1 h 49 min | 177 | 206 | $12.08 |
| Animate Act II: Gold into Lead | 18:24–20:36 | 2 h 12 min | 177 | 227 | $14.26 |
| Animate Act V part B | 19:29–21:13 | 1 h 44 min | 179 | 197 | $14.51 |
| 20 agents | 2,420 | 3,335 | $206.21 | ||
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. 4 h 48 min in all.
Each ran from its first brief to its handback, with a polish pass on its own act after the director’s review.
50 plates from 17:55 to 20:58 and 141 sprites, with 131 voice lines and 27 ambience and effects clips. Backcloths and cut-out pieces are generated; all motion, light, weather and type are drawn in code.
A toy-theatre engine adapted from The Long Reach: flats slide on from the wings, backcloths drop from the flies, and dramatic moments turn the figures into black silhouettes. Palatino type, with TeX Gyre Pagella embedded as a fallback.
A Scottish bed composed in code (reed drone and chanter, clarsach, fiddle, drums, a recurring crown motif, traditional airs from ABC) under Kokoro bf_emma at speed 0.9, with three character voices for attributed quotations.
1376 by 768 at 30 fps, 6.7 Mbit/s with AAC audio; the first export wrote 34,710 frames in 45 minutes. The single-file page is 124 MB.
The director read each act from a whole-act contact sheet and sent a numbered polish list to its own animator; the fact check read five claims against primary texts.
| Found | What went wrong | How it was fixed |
|---|---|---|
| Fact check, John Damian | The excuse for his failed flight from Stirling was paraphrased loosely, and 1507 is the modern date, not Lesley’s | The line now has him fall and break his thigh, blaming hens’ feathers that longed for the midden, not the sky, as Lesley tells it |
| Fact check, Carey’s ride | “Galloped” overstated it, and his memoir gives no total duration | The narration says a rider spent three days in the saddle riding north to tell James he was King of England as well |
| Fact check, James V’s last words | “It cam wi’ a lass” is not in Knox, who has “it came from a woman; and it will end in a woman” | The quotation follows Knox: “It came from a woman, and it will end in a woman” |
| Fact check, the Coronation Chair | It was made to hold the Stone (about 1297 to 1300), but it was not Edward I’s own coronation chair | The narration says Edward had a coronation chair made to hold the Stone |
| Flat and ghostly pieces | Flags drawn as flat filled shapes, Picts that read as “penny plain” ghosts, a flat bright-green mound at Scone, genealogy words that read as scratches | Flags engraved, the Picts repainted, the mound and the genealogy redrawn; rule 3 of the director’s notes came out of it |
| Black frames and wrong depth | Black frames between scenes, and pieces hidden behind a backcloth’s near layer | Scene changes made by stage machinery, and a depth rule written into the director’s notes |
| Sound levels | A ten-gun salvo, battles and 50 hammer blows set the bed’s peak 10 dB over the music, so the music dropped and the mixer would have lowered the whole film | One-shot clips trimmed to their first strike and the effects bus soft-limited at −3 dB, 6 dB under the music |
| Left in | A few first-draw hitches of under a second when a shot’s cards are built, and a code-drawn Langside cloth | Listed as known small things in the director’s notes |
Nine Sonnet animators cost $108.18, 52% of the $206.21. Act I: The Stone was the dearest at $17.56, and Act II: Gold into Lead ran longest at 2 h 12 min.
Nine Haiku researchers cost $5.86 between them and the last reported at 17:54, before the first animator started at 18:24.
63% of the cost was cache reads, the long conversations read again on every call. The director alone reached a 966K-token context and was summarised once, at 18:48.
53 min of the director’s hours went on waiting for agents, against 1 h 05 min on engine and film code, 35 min on score and sound and 34 min on plates.
You said at 18:42 that the narrator was fine, that no voice comparison was needed and that bf_emma should stay, which saved a round of auditions.
The film was reviewed from act sheets and frame captures, and a full scan of 4,628 frames found no errors, but no agent has heard it. You accepted the first five minutes of sound on an audio-only listen.
Measured from the session's own transcripts: the director's, and one for each of the 19 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 ran in the crown-scotland-showcase worktree and continued once after its context was summarised; both halves are in the same transcript. Gaps of more than 20 minutes (41.7 minutes in all) are left out of the hour-by-hour split.
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 findings come from the project’s POLISH.md, DIRECTOR_NOTES.md and research/10_fact_check.md, and the film figures from the lock file and ffprobe of the published MP4. 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.