A fifteen-minute film about Marie Tharp, who drew the first picture of the rift valley in the Atlantic seafloor from soundings she was not allowed to collect, told on a drafting table where black ink is what a ship measured, blue is what she inferred and red is what is at stake. Sonnet directed a fleet of Haiku and Sonnet agents, with no Opus, and delegated all the coding. 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 four local rows mark the pictures, voice lines and audio files the machine produced. Eight checkers went first, five pilot agents followed, the engine integrator and composer ran through the middle, and the ten chapter animators worked in two waves.
| Agent and assignment | Ran (local) | Time | API calls | Tool calls | Cost at list prices |
|---|---|---|---|---|---|
| Director · Sonnet 5.5 | |||||
| Director: orchestration, review and export | 06:46–14:30 | 7 h 44 min | 282 | 361 | $36.51 |
| Research and checking · Haiku 5.5 (one Sonnet) | |||||
| Fact-check: early life and Lamont hiring | 06:50–06:59 | 10 min | 18 | 39 | $0.20 |
| Fact-check: rift discovery and argument | 06:50–07:04 | 14 min | 54 | 87 | $1.56 |
| Raw-text check of Tharp essay, 1959 paper | 06:59–07:09 | 10 min | 29 | 36 | $0.34 |
| Fact-check: women barred from ships | 06:50–06:58 | 8 min | 14 | 36 | $0.13 |
| Period reference: 1950s Lamont lab, tools | 06:50–07:00 | 10 min | 13 | 42 | $0.18 |
| Fact-check: maps, Berann, honours, legacy | 06:50–06:58 | 8 min | 16 | 38 | $0.18 |
| Verify Tharp quotes and misattributions | 06:50–06:59 | 9 min | 16 | 41 | $0.09 |
| Fact-check: data, profiles, Navy limits | 06:50–06:58 | 8 min | 12 | 36 | $0.16 |
| Look pilot · Sonnet 5.5 | |||||
| P5: world ridge system, grid, world sheet | 09:37–10:25 | 48 min | 129 | 160 | $7.73 |
| Pilot P4: look test plates and hand study | 07:19–12:36 | 5 h 17 min | 402 | 594 | $33.82 |
| Pilot P3: geo data, relief, flight | 07:19–09:43 | 2 h 24 min | 325 | 374 | $27.66 |
| Pilot P1: ink engine, six profiles | 07:19–13:36 | 6 h 18 min | 489 | 558 | $45.02 |
| Pilot P2: light table, the dots shot | 07:19–09:34 | 2 h 15 min | 252 | 344 | $22.77 |
| Helpers · Haiku 5.5 | |||||
| Haiku: production plates fragment with provenance | 10:46–12:37 | 1 h 51 min | 19 | 23 | $0.42 |
| Voice audition: Kokoro narrator voices | 07:19–07:25 | 6 min | 17 | 24 | $0.11 |
| Engine freeze · Sonnet 5.5 | |||||
| E1: engine freeze and integration | 09:44–14:30 | 4 h 45 min | 564 | 621 | $52.45 |
| Score · Sonnet 5.5 | |||||
| C1: composer, plan and requests for the score | 09:48–14:14 | 4 h 26 min | 435 | 480 | $39.12 |
| Animation, one agent per chapter · Sonnet 5.5 | |||||
| Animator ch6: The dots | 10:43–11:59 | 1 h 16 min | 208 | 250 | $13.93 |
| Animator ch8 plus s7_5: Kane and the sled | 11:58–13:06 | 1 h 09 min | 234 | 254 | $16.43 |
| Animator ch9: The painting | 10:43–13:18 | 2 h 36 min | 364 | 386 | $35.12 |
| Animator ch3: Girls were needed | 11:58–12:55 | 57 min | 217 | 225 | $12.38 |
| Animator ch2: Papa's maps | 11:58–13:24 | 1 h 27 min | 247 | 281 | $19.68 |
| Animator ch7: The gully (five shots) | 11:58–13:33 | 1 h 36 min | 266 | 312 | $19.55 |
| Animator ch1: The legend | 10:43–11:27 | 44 min | 148 | 163 | $8.79 |
| Animator ch5: Six profiles | 10:43–12:21 | 1 h 38 min | 201 | 223 | $18.05 |
| Animator ch10: After, and the credits | 11:58–13:28 | 1 h 30 min | 268 | 294 | $20.69 |
| Animator ch4: Lamont Hall | 11:58–13:18 | 1 h 21 min | 231 | 282 | $20.11 |
| 28 agents | 5,470 | 6,564 | $453.17 | ||
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. 6 h 59 min in all.
Chapter 9, the painting, took the longest, and its last frame had to be a painted panorama pulled well back.
Production plates between 10:48 and 12:37, after 155 pilot pictures from 07:21. The relief is drawn from NOAA’s ETOPO 2022 bedrock grid (491 MB) and the ridge from the PB2002 plate boundaries, both downloaded with your approval.
bm_fable narrates 34 lines and af_heart reads Tharp’s own written words in 11, 1,591 words in all, both chosen by you from the audition. The voice lines were made between 09:46 and 11:56.
YuE2 piano and cello, cut and arranged. Dense pieces followed the lead sheets, sparse ones failed, so phrases came from real takes. Effects are generated with Stable Audio and sit about 17 dB under the music.
1920 by 1080 at 30 fps. Frames were checked from the encoded file, with no black frames at any start, end or chunk boundary. The 130.6 MB single-file page matched the unpacked page to within 1.46 of 255 with no console errors.
Frames, strips and measurements stood in for a real-time viewing, and these are the failures they recorded, with how each was handled.
| Found | What went wrong | How it was handled |
|---|---|---|
| Six profiles (pilot) | Latitude labels stated something the data did not support | Removed; the profile view was pushed in because it was too small |
| Light table (pilot) | The mechanic worked but the palette was amber and the props flat | A numbered polish pass, and prop sprites made by the look agent |
| Relief (pilot) | It looked like default computer graphics | Polished, and the look agent’s table plate adopted |
| Fetch tools | They summarise pages, so a quote could not be trusted | Quotes checked from raw text or page images, by a clerk and a quotes verifier |
| GPU use | An unconditional comfy.py down killed another agent’s plates job | A GPU guard and lock file; never bring ComfyUI down while it is busy |
| YuE2 music | Sparse pieces failed, because the rests were not quiet | Phrases cut from real takes and the silences gated in the arrangement |
| Smoke test | Reported a hang in one shot | A Node stub artefact; the browser measured a median of 22 ms |
| Left in | Speckled start on the town plate’s reveal, one faint ghost frame, the globe’s brass ring cropped in one frame, an unused drawer hand | Accepted as polish enough, and listed in the hand-over |
The pasted outline came from an earlier session and was treated as untrusted. Eight checkers verified it before any storyboard was written, and the results went into research/FACTS.md, which every agent had to use.
| Point in the pasted outline | What the research found | In the film |
|---|---|---|
| Who made the epicentre dots | Howard Foster, a deaf draughtsman, plotted them; Bruce Heezen noticed the line they made | The credit line is stated precisely |
| What convinced Heezen | The East African rift, in mid-1953 | Follows the research |
| Cousteau’s film | 1959 | Follows the research |
| Maurice Ewing | Interested in the valley, and opposed to continental drift | Follows the research |
| Whether she went to sea | She did, on the Navy ship Kane in 1968; a 1965 sailing is disputed | Follows the research |
| “The four greatest cartographers” | Unsupported | Not used |
| The 1977 map | The single “first comprehensive map”, published by the Office of Naval Research | Follows the research |
The Sonnet-directed fleet cost $453.17 for 15.1 finished minutes, $29.99 a minute, against $7.56 to $13.18 for the other films built by agent fleets. Every agent ran on Sonnet or Haiku, as the brief asked. The price came from long conversations, not from the model rate.
Five pilot agents cost $137.00, 30% of the total, before a shot of the film proper existed, and the longest ran 6 h 18 min. The pilot is where the look, the ink engine, the relief and the flight were settled, and it is why the chapters then ran in parallel.
5 h 14 min of the director’s 7 h 44 min went on waiting for agents, and only 5 min on code, because all coding was delegated. 22 min went on briefing and messaging them.
6 agents reached contexts of more than 900,000 tokens (the engine integrator, the composer, the first pilot agent and the chapter 7 and 9 animators) and cost $210.93 together, 47% of the bill. A fresh agent for each round of work is the obvious lever.
The ten chapter animators cost $184.74 (41%), ran 44 min to 2 h 36 min, and started in two waves, at 10:43 and 11:58. ch9: The painting took longest.
You said the clips sounded okay and the reel sounded good. The full mix, the narration end to end, the effects and the web film’s audio have not been heard by anyone, and nobody has watched the film in real time.
Measured from the session's own transcripts: the director's, and one for each of the 27 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. No Opus agent appears in the transcripts; every model is Sonnet 5.5 or Haiku 5.5.
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 corrections to the outline, the pilot findings, the audio decisions and the left-in items come from the project notes and the director’s final report. Export figures come from the project files. 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.