Back to Token Sink
Real-time film · 9:11 · production notes · 17 October 2026

Footfall

Footfall, 9:11. Watch on YouTube.

A nine-minute film that follows one unbroken trail of tracks from a worm-burrow on a Precambrian seabed to bootprints on the Moon and rover treads on Mars, told as light shining up through a black stage floor in coloured shadow theatre. Nineteen agents built it in under three hours. This is where the time and the money went.

The session in four numbers

From a request to a published film

2 h 45 minfrom the request at 10:10 to the commit at 12:54
14 h 29 minof agent working time, run in parallel
19 agentsone Opus director, 14 Sonnet and 4 Haiku agents
$148.37if every call had been billed at API list prices
9:11 film1920×1080 at 3045 shots50 lines, 1,430 words1 voice48 plates18 sprites1,351 sound cues11,600 lines of film code
The afternoon

10:10 to 12:54

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, sprite and voice line the machine produced. The four researchers went first, the director ran the pilot and wrote the script, the engine builders came next, and eleven chapter animators worked in one wave.

10:0011:0012:0013:00Directorresearch: deep time tracksresearch: human origins migrationresearch: civilisations and explorationresearch: industry, space, computingpuppet enginesprite rig engineComposerc04 Uprightc10 Invisible trailc01 First trailsc02 Onto landc03 Giantsc05 Spreading outc06 Settlingc07 Over the horizonc08 Ironc09 Leavingc11 Next (shots 1-3)Local: plates (48)Local: sprites (18)Local: voice lines (50)audio approved 11:46MP4 written 12:48
Opus 5.5Sonnet 5.5Haiku 5.5Local machine, no API cost
The company

Who did what, for how long, at what price

Agent and assignmentRan (local)TimeAPI callsTool callsCost at list prices
Director · Opus 5.5
Director: orchestration, review and export10:10–12:542 h 45 min288400$46.14
Research · Haiku 5.5
Research deep time tracks10:18–10:212 min1431$0.04
Research industry, space, computing10:18–10:212 min917$0.03
Research civilisations and exploration10:18–10:202 min1126$0.02
Research human origins migration10:18–10:213 min616$0.05
Engines · Sonnet 5.5
Puppet engine: human rigs10:50–11:1525 min6895$3.04
Sprite rig engine10:50–12:331 h 42 min304400$19.19
Score · Sonnet 5.5
Composer: YuE2 score11:23–12:411 h 18 min158163$7.15
Animation, one agent per chapter · Sonnet 5.5
Animator c01 First trails11:26–12:0034 min85129$4.66
Animator c10 Invisible trail11:23–12:1552 min104137$6.63
Animator c09 Leaving11:35–12:2650 min130160$8.75
Animator c11 Next (shots 1-3)11:35–12:0832 min99116$4.91
Animator c07 Over the horizon11:35–12:1439 min86102$4.92
Animator c03 Giants11:35–12:1337 min97140$5.34
Animator c08 Iron11:35–12:381 h 02 min167213$9.52
Animator c02 Onto land11:35–12:1640 min119146$6.02
Animator c06 Settling11:35–12:2650 min151183$8.30
Animator c04 Upright11:23–12:0037 min116143$5.95
Animator c05 Spreading out11:35–12:2954 min136161$7.69
19 agents2,1482,778$148.37
The bill

Where $148.37 went

By role

Director · $46.14 · 31%Pitch, pilot, script, review of every chapter at full size, export and publishing.
4 Haiku researchers · $0.14 · 0.1%Short research notes, capped at ten minutes and twelve fetches each.
2 engine builders · $22.24 · 15%Jointed human puppets, and sprite rigs with planted feet.
Composer · $7.15 · 5%Wrote its own ABC scores for YuE2 and aligned the renders.
11 chapter animators · $72.71 · 49%Eleven Sonnet agents in one wave.

By model

Opus 5.5 · 1 agent$46.14 · 31%
Sonnet 5.5 · 14 agents$102.10 · 69%
Haiku 5.5 · 4 agents$0.14 · 0.1%

By kind of token

Cache reads (802 million tokens)$93.94
Cache writes (12 million tokens)$37.81
Output, thinking included (1.4 million tokens)$16.57

The director, hour by hour

waiting on agents1 h 13 min · 44%
engine and film code29 min · 18%
plates and images (generation)15 min · 9%
review and checking11 min · 6%
score and sound9 min · 5%
export, video and publishing8 min · 5%
agents: brief, launch, message6 min · 4%
docs, memory and notes6 min · 4%
other and unclassified3 min · 2%
waiting on the user3 min · 2%
research, script and brief2 min · 1%
clean-up1 min · 1%

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. 2 h 45 min in all.

The animators

How long each chapter took

Every chapter was reviewed at full size and sent one or two polish rounds.

Eleven chapters, one animator each

c08 Iron1 h 02 min
c05 Spreading out54 min
c10 Invisible trail52 min
c06 Settling50 min
c09 Leaving50 min
c02 Onto land40 min
c07 Over the horizon39 min
c04 Upright37 min
c03 Giants37 min
c01 First trails34 min
c11 Next (shots 1-3)32 min
The film

What the afternoon produced

Nine frames from the film: a blank card theatre with the title, a turquoise Cambrian sea over a punched trail of trilobite tracks, a sauropod against an orange sky above rows of footprints, an orange map of Africa with a dotted migration path, a canoe under the Milky Way, a horse on a wagon against an orange sky, an astronaut at the Apollo lunar module, the Curiosity rover against a Martian dusk, and the trail as a line of light on a dark blue field.
Nine frames from 12 s to 540 s: the title, the Cambrian, the dinosaurs, the migration, the canoe, the horse, the Moon, Mars and the final trail.

The local machine

48 plates, 18 sprites

Plates from 10:24 to 11:55 and sprites from 11:04 to 12:24, each cut into layers for the theatre. The voice lines took about a minute of CPU time at 10:49.

One voice

50 lines

1,430 words read by bm_fable at speed 0.95, as you chose. You heard and approved the opening audio at 11:46 (“sounds good, nice mix”); the rest of the mix was measured only.

The score

YuE2, arranged

The composer wrote its own ABC scores for YuE2, whose planner ignores key and tempo, and aligned the renders with chroma DTW and a time-stretch. Footfall percussion comes from the step events.

The MP4

9:11, 232 MB

1920 by 1080 at 30 fps, exported in parallel chunks, with a 28.7 MB single-file page and a 540p share copy. The browser scan drew 552 frames, one per second, with no console errors.

What the checks caught

Eleven findings and what was done

The review log records the pilot draft’s faults as found, and the recurring problems the director briefed up front for the later chapters.

FoundWhat went wrongHow it was handled
Pilot draft: stagingThe horizon sat about 100 px above the ground line, so the walkers’ legs vanished against a dark plainPlates rescaled so the walkers stand against the lit hills
Pilot draft: the figureA rectangular torso slab, arms lost inside the silhouette, a boxy hood, a crude fur fringe and joint pins that read as dustPuppet rebuilt on a test bench, with slimmer torso, separated limbs and pins removed
Pilot draft: the printsHominin prints read as tadpoles, and older tracks were as bright as the hero trail, a dotted highwayPrints redrawn as heel, arch, ball and five toes; older tracks dimmed so the fresh trail leads
Pilot draft: the cardA hard white bar where the light leaked over the card edge, and black grass invisible on blackThe light confined to the card; tall grass rising into the light
Smoke and cloudsA cartoon chain of circlesEvery later brief asked for one-fill scalloped lace-cut lobes
Generated peopleThey lose limbsPeople drawn as code puppets
Glowing slabRead as a modern tablet computerRedrawn
LabelsPushed off the bottom of the frameRepositioned in review
Weak spritesThe Voyager and the first-pass horse were weakBetter sprites generated, and the galloping horse rebuilt as a rigged sprite
Smoke testReported an errorA fault in the Node test harness; a browser scan of 552 frames was clean
Left openA theropod hip notch, roundish launch smoke, a grass blade crossing the “Rainhill, 1829” label, and a horse that only tosses its head instead of looking backAccepted as minor and listed in the notes
What it taught

Six things the numbers say

Method

How these figures were measured

Measured from the session's own transcripts: the director's, and one for each of the 18 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. Figures run to the commit at 12:54.

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 pilot findings come from the project’s review log and the open items from the director’s notes. The file sizes and dates 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.