One More Take

The crew, exactly

This page is the honest version. Every claim below comes from the coordinator's timestamped notes, the builders' own logs, or the councils' saved output. Where something was not checked, it says so.

Not made by agents: the car, its handling and the vehicle template content are Epic's, shipped with the engine and used untouched. The real sound recordings and the music tracks come from the jam sponsor WeLoveIndies. The engine's own Roboto and Droid Sans Mono fonts draw the HUD.

The brief

At 11:09 on Sunday the director handed the jam page to the coordinator with one instruction, in short: enter the jam, use the sponsors' tools, plan with a council of agents, and build with Opus 5.5 workflows.

By 11:20 the rules were in hand (Epic's own news page refused automated fetches, so they were read from the jam's itch.io page and Epic's forum thread). Two facts reset the day. The theme was DIRECTOR'S CUT. And the deadline was that night, 01:59 GMT Monday, not a day away: about fourteen and a half hours, with a Windows build required.

• 11:10Machine survey on the Linux workstation: no Unreal, no Blender, no admin rights.
• 11:23Blockers found: the engine download needs an Epic sign-in, there is no jam registration yet, and the Windows build box is unreachable.
• 11:31The director signs in to Epic, joins the jam, registers with WeLoveIndies and rules out a mobile build. The engine download starts (a 42 GB archive).

No engine, no admin, no problem (mostly)

Blender 5.3 was installed as a user package and ran headless at 11:27. The Linux engine unpacked to 77 GB at 11:50. By 11:52 there was proof: a blank C++ project compiled in 75 seconds, the editor ran offscreen on Vulkan, and the engine's built-in MCP server answered. That is the shape of the whole day: the agents never saw a window. They drove the editor from Python, played the game through injected key presses, and judged it from screenshots they opened themselves.

The tooling the builders wrote before and around the game is most of what made a crew of agents workable:

• One compile at a time. A build script holding an exclusive lock, so eight agents in one source tree never compile over each other.
• Autoplay. The game drives itself through injected input, takes screenshots, and writes a result.txt with a pass=N fail=M line and an exit code. A bot with Perfect, Sloppy and Lazy driver profiles came later.
• Capture. A frame-dump route to a 1080p H.264 video, with the sound rebuilt from the game's cue log.
• Windows scripts. Sync, package, zip and smoke test on the Windows build box over the network, checked against a fake engine because the real one was not there.

The Linux engine has no Windows binaries, so the Windows package has to be built on the Windows build box. That one fact turned into the longest-running problem of the day (Scene 4).

The council that picked a puzzle

While the engine downloaded, the coordinator convened nine Fable agents: four designers, two judges, one architect and two critics. They ran for 29 minutes and produced eight concepts.

The judges disagreed. One preferred a reflex runner, CUT TO THE CHASE; the other a turn-based puzzle, CUTTING ROOM FLOOR. The architect took the puzzle on combined score (84 of 200 against 76) and named it FINAL SPLICE: a film strip is the level, the player moves, cuts and restores shots, and a clapperboard stunt star hops through them in the order spliced. The spec's own reasoning was that a turn-based game can be proven solvable by a solver before the engine exists, and nothing in it depends on the feel tuning that agents cannot judge.

The two critics returned three blockers and about thirty findings. At 12:00 the coordinator wrote rulings: a third reel, gamepad, mouse, the pause menu and a Blueprint fallback were cut; every tier kept all eight scenes; the Windows packaging proof moved to the front.

An 80-minute first playable

Ten Opus 5.5 builders started at 12:01. Rules and levels, meshes, audio and the Windows scripts ran from the first minute; the Unreal skeleton with frozen interfaces landed at 12:17; game logic, HUD with audio, the stage and content tooling then ran in parallel in one source tree, sharing the build lock.

At 13:21, eighty minutes and about two million subagent tokens later, the puzzle was playable end to end:

Integration fixed three seam bugs no single builder owned: a map script that only worked on its first run, take shots blown out to white by a glowing selection rail, and engine debug text drawn over the HUD.

Meanwhile, on the Windows build box

From 12:10 the Windows side stalled. The Epic launcher showed a greyed-out button where engine versions are added. Hours of the director's time went there. The coordinator diagnosed it from the launcher's logs and an open-source Epic client: the account's library held no Unreal Engine entitlement at all. The Windows scripts polled for the engine every five minutes for an hour and never saw it. At the time this page was written, it was still not installed.

Five minutes, four designers, one stunt

First, an inventory of what ships with the engine that agents cannot make well themselves: Epic's tuned Chaos sports car and offroad car, a time-trial track, mannequins with locomotion, Niagara explosion templates, fluid fire and smoke, volumetrics, Chaos destruction. The new bar, written into the brief: within ten seconds a judge thinks "that is an Unreal game", and within thirty they are having fun without reading anything.

Then a second, faster council: four Fable designers, one lens each, one concept each, five minutes. Every one of them landed on a stunt shoot on a film backlot.

At 13:50 the coordinator merged 1 and 3 into ONE MORE TAKE and wrote the brief. The theme became three verbs the player does: shoot what the studio cut (red-flagged stunts signed CUT FROM SCRIPT), roll the hero camera (time drops to 0.3x, a film meter burns, stunts on film pay triple), and call the final cut (PRINT IT or ONE MORE TAKE).

What survived the cut from FINAL SPLICE: the project skeleton, the ink-stamp HUD, the audio manager and its 56 cues, the autoplay harness, capture and packaging. The puzzle itself was retired, read-only.

The rebuild on Epic's vehicle template

Three builders went first: a foundation, the meshes and the synthesised audio. Then six more ran in one tree: director and scoring, set pieces and lighting, hero camera and replay, HUD and audio, scenes and driving bot, content pipeline and packaging.

• 13:55Sponsor audio: a WeLoveIndies chest of 24 sounds and 5 music tracks under the Free & Demo licence, at no cost. Checkout stopped at the licence-owner form: an address is the human's to give, not the agent's to guess.
• 14:0541 synthesised driving cues, including four level-matched engine loops built from simulated V8 firing pulses, and four original music cues. Checked by spectrogram and measurement. Nobody had listened.
• ~14:08First six film-set props out of Blender. By the end: 30 props (36 meshes counting the breakaway billboard pieces), 63,704 triangles, six tileable texture sets, and a contact sheet with a 1.8 m bar beside every mesh so scale can be checked by eye.
• 14:25Foundation lands: a new project from Epic's vehicle template, frozen interfaces, scene 1 end to end under the driving bot, exit 0. Ten minutes ahead of its gate. A first editor run had stalled for ten minutes on a full shader recompile triggered by the template's renderer settings; those three settings were put back.
• 14:25–15:05Six builders in parallel. 112 assets imported with 0 errors; 25 materials; collision proven on the ramps by 400 line traces; a scene built at runtime from data in about 0.1 s with 220 physics bodies; a smashed set reset in about 4 ms.

One finding shaped everything after it: the car is not slow. Flat out it reaches 158 km/h; the kicker ramp at 139 km/h throws it 52 m and it lands on its wheels. Epic's car is very stable in the air. That is why holding W turned out to be a strategy.

A deadlock nobody wrote

The Shipping compile inside the Linux packaging run hung: the compiler was blocked writing to a pipe the build tool had stopped reading, and it was holding the shared build lock. Three builders queued behind it for twenty minutes. Two of them diagnosed it down to the process id and reported it instead of killing another owner's job, which is what their rules said to do. The coordinator killed it at 15:29 and the queue drained.

The second lesson came a few minutes later. At 15:36 about twenty game processes were running at once on one GPU. Frames took 131 to 250 ms, the bot crashed every take for reasons that were not real, and every measurement from that window was worthless. The fix round runs games through a two-slot queue and discards any run averaging above 25 ms a frame.

Four critics: "he would reject it"

At 15:30 the coordinator went through the screenshots of the rebuild. It read as an Unreal game now: the car, forty-odd crates in the air, a hero-camera freeze of the jump, a production-report result card with both stamps. Faults sent back: the driving camera sat so low the car hid the road, one replay angle was pure motion blur, the night preset was a fog white-out, and golden hour was muddy.

The first playable was committed at 15:55. Four critic agents, each with one lens, reviewed it with a producer agent turning their findings into ranked work orders. Asked whether the director would reject it as it stood, all four said yes.

"Holding W and braking at the end earns THEATRICAL CUT and $93k (CULT CLASSIC is $90k per scene), while the Perfect bot has not produced a single DIRECTOR'S CUT in any scene."Critic, game feel

"Three of the four scenes are so underexposed they read as night (mean screen brightness 8-24%, no pixel above 78% in any daytime still), and the scene-1 pyro is an orange fog blob, not a fireball."Critic, look

"A first-timer who overshoots the mark or runs below about 25 fps is wrecked and loses the whole take without ever seeing a result card."Critic, first-time player

"There is no Windows build and no way to make one yet."Critic, ship and performance

The critics also listed what must not break: the hero-camera frames, the paperwork HUD, launch to driving in under five seconds, wreck to ACTION again in 1.7 seconds with lines like "WRONG SIDE UP. AGAIN.", Epic's car untouched, and scene 1 at 11.8 ms a frame in the Shipping package on a quiet GPU.

Their honest scores out of 5 sat mostly at 3, with use of theme at 4 and visuals at 2 from the one critic who measured the pixels.

One more take

The producer's headline had three points, in order:

1. A first-timer must always reach a result card and the car must always answer the keys. An overshoot, a stop short or a time-up was a wreck paying $0; a held W went dead after the slate; below 30 fps the car slowed while the take clock did not.
2. DIRECTOR'S CUT must be earnable, and holding W must not match it. In the 15:41 sweep only scene 2 produced one. Scene 3's red flag could not be landed at all; scene 4's fire ring was missed by 7 m.
3. Daytime must look like daytime and explosions must look like fire. Brightness targets measured on screenshots, not judged by mood.

Eight builders took the orders, one owner each, ranked, with acceptance checks and a 50-minute time-box, followed by one integrator and an acceptance checker told to refute the claim that the blockers were fixed. The car's handling stays exactly as Epic shipped it; difficulty is tuned through the scoring table and the levels, not the vehicle. A play session by the director was pencilled in for about 17:30.

Round two, in numbers

Done at 17:05: 10 agents, 71 minutes, about 2.0M tokens.

The list: the scene 1 crate wall could launch a flat-out car, the scene 2 breakaway facade nearly stopped it dead, explosions were white puffs, and the world was a flat sand plane with blotchy tarmac. The coordinator opened the stills and agreed: daylight at last, and the hero-camera frames looked like a film, but the driving view was still plain. Still no Windows engine. The 17:30 play did not happen; the director was not heard from between 15:45 and the evening, and the crew carried on without waiting.

A second lock nobody wrote, and a judge who says 41

Round three went out at 17:10 with eight builders (breakaway physics and pyro, the world's look, cameras and replay, game flow, scenes, HUD and audio, tooling, and a builder for the gameplay video), then an integrator, an adversarial acceptance check and a new role: a first-minute judge who sees only what a jam judge would see, the stills and the video, and scores against the jam's criteria.

• 17:14–17:27The second deadlock. An editor run left two engine helper daemons (a trace server and a cache server) alive, holding the content lock's file handle. Three content jobs queued behind nothing for 12 minutes. A builder traced it to the process ids and reported it rather than kill what it had not started; the coordinator checked the process table, stopped both, and ran a watchdog for the jobs already queued under the old script.
• 18:25Done: 11 agents, 78 minutes, about 2.5M tokens. The scene 1 crate wall no longer flips a flat-out car. The "breakaway" facade had cost 85% of the car's speed because of its solid posts, not its panels; now 0.1%. Red flags need the real jump. The road blotches were a wet-asphalt material full of puddle shapes; the world got worn tarmac, textured ground, numbered sound stages, a perimeter wall, palms, hills and volumetric clouds. Acceptance: go.

"Four straight roads that each open on the same crate wall."First-minute judge, round three: 41 of 70, would reject again, narrowly

The judge gave theme an 8 and liked the hero camera, but listed pyro that was not there, a bot that won everything first time, and a video still showing the old world.

Rules that push back

Round four took the judge's list: scenes that differ, a clock that pushes back, a Director's Cut that needs every red flag on film, real fire and smoke, a photograph of the take's best filmed moment on the result card, and a video pipeline that cannot lose footage.

The four scenes now open on different breakaways (crate wall, fruit stall, water barrels, picket fence) under four different lights: golden hour, wet overcast, pink dusk, and night with moon and stars. The result card carries "the print", a photograph the hero camera took of your best moment, taped to the daily production report, and the end screen lays the prints out as a film strip.

"No, I do not think the director rejects it again on looks, but it is close."First-minute judge, round four

At 19:20 the director took the play lock, the switch that means a human is at the controls, and held it for seven minutes; every agent's build and run waited, and nobody touched the lock. The round-four package was committed at 19:27 and kept as the fallback for the final round.

What the checker and the judge still found: night explosions with no fire, smeared hero frames in scenes 1 and 2, the car turning yellow behind the result cards, the rank stamp covering two prints on the end screen, a scene 1 glass red flag that a pyro charge breaks for free on the safe line, and four scenes that are still straight strips of road.

What it looks like now

Every still below was taken by the autoplay bot in the frozen round-five build: the integrator's four-scene run at 20:50, plus the two pyro stills from the pyro builder's test runs of the same code (the title comes from round four; the title screen did not change). Each was opened by an agent before it was used here. Earlier stills are on this page where the story needs them, with their times.

The honest part

What went wrong

• The first game was cut. Eighty minutes of ten agents, 121 passing checks, and the director rejected it in one look. Tests measured correctness, not whether it was any good.
• The Windows engine install blocked for hours. The account had no engine entitlement and the launcher would not offer an install. Every Windows script was written and tested against a stand-in.
• The council's pick was safe for agents, not good for players. The reasons the puzzle won (solver-provable, no feel tuning) are the same reasons it was easy.
• Too many games on one GPU. At 15:36 about twenty game processes shared the card and every timing from that window was discarded, including one critic's whole sweep.
• Two locks nobody wrote. A hung compile held the build lock for twenty minutes (15:07); two orphaned engine daemons held the content lock for twelve (17:14). Both times builders found the process ids and reported them instead of killing another owner's job.
• Polish judged by agents, not people. From 15:45 to the evening the crew worked without the director's notes. The "first-minute judge" is an agent reading stills; its 41 and 48 are estimates, not a jury.
• Collision tools that silently did nothing. Under the editor's script commandlet the collision setters returned success and made no shapes; convex decomposition returned zero shapes even in the full editor. Every physics prop got one box instead.
• A shader trap. Three renderer settings inherited from the template forced a ten-minute global shader recompile before the editor would open.
• The car is quick and stable, so at 15:41 a bot holding W out-scored a careful one; the fix round answers that in the scoring table and the levels, not the car.
• A 100-second frame at the hand-over from the last result card to the end screen, found at 14:25, traced to the render thread; autoplay passed regardless.

What is still weak or unproven

• Pyro by day. The white bleach is gone, but at 0.4 s the fireball reads as brown-orange dust, not flame. An additive flame on bright sand either clips white or vanishes; fixing it needs a material change, which the polish round did not allow. At night there is now an orange fireball, but the scene 4 FIREBALL hero frame shows the ring and glow, not distinct flame tongues.
• The car changes colour. Behind the result card it reads yellow, and at night its top reads orange-yellow where a beam hits it. Exposure was ruled out; the cause was not found.
• A free red flag. In scene 1 a pyro charge can break the glass red flag for a careful driver who never jumps. Moving the charge made it worse, so the data was put back; fixing it needs a code change.
• No buses in the scene 3 hero frame. The jump is real and filmed; the camera picks its shot before the car is over them.
• Money order. Scene 1 still pays more than scenes 2 and 3; global payout rates cannot fix that without scene changes.
• Straight roads. All four scenes are straight strips of 420 to 540 m. They differ in light, set dressing and stunts, not in shape.
• Nobody has listened to the audio end to end. Cues were checked by spectrogram, level measurement and the game's cue log. The engine sound is four synthesised loops. No one has heard a whole run with the mix as shipped.
• No Windows build. The Windows build box never got Unreal Engine 5.9; at the last check (20:49) it was still not installed, so there is no Windows package. The Windows scripts were tested against a stand-in. The game ships for Linux only.
• The dailies replay is off in the shipped build. It was built, but the replay hid the car too often, so the result card carries the print instead.
• Easy for a good driver. The Perfect bot takes every Director's Cut first time with 8 to 11 seconds to spare. A human-like bot found scene 3 hard in round four (4 of 10); after the far ramp was widened it took 6 of 8.
• Physical controller input has not been tested on hardware; gamepad support rides on the template's own input mappings.

Agent runs

Throughout, one coordinator agent (Claude Code running Fable 5.2) and one human. The coordinator commits; builders never run git commands that change anything, never publish and never sign up for anything.

Final hours

Round five started at 20:30 with one rule: polish only. Numbers, strings, colours, layout, camera placement and documents; no new systems, no change to how a scene is completed, every change revertible in one edit, and any change in behaviour kept behind its existing switch. The round-four package stays on the shelf as the fallback and ships if this round breaks anything, so a broken scene means that builder's work is thrown away.

• 20:30Thirteen agents in the round: nine builders on the judge's and the checker's lists, then the integrator and the final documents.
• 20:28HUD. The end-screen rank stamp moves off the film strip to sit beside the total. Speedometer digits shrink 15% above 100 km/h so they clear the dial labels. The in-drive prompt now says how: "LAND IT ON FILM (HOLD SPACE) FOR THE DIRECTOR'S CUT".
• 20:32Pyro. The day mortar flash drops to a tenth of its peak (30000 to 3000 cd) with half the radius; pixels over 90% white fall from 6.2% to at most 1.1%. Night flames go from 2.6 to 8 and the scene 4 FIREBALL burst from 0.75 to 6.5: an orange fireball at night where round four showed nothing. Brighter market awning, a light-blue glass pane on a yellow rig.
• 20:35Rules and money. A flip now blames a set piece only if the car hit it before turning over, so the forced-wreck card reads "SHINY SIDE UP NEXT TIME." Payouts moved, no rules: air pays more (600 to 1000 a metre), smashes and near misses less, red flags 40000 doubled. Scene 4's share of a perfect run fell from 41% to 29.5%; holding W totals about $245k, STRAIGHT TO VIDEO. Scenes 2 to 4 get 30 s clocks; scene 3's far ramp widened from 4 m to 5.5 m (a human-like bot's Director's Cut there went from 4 of 10 to 6 of 8).
• 20:40World. Three in four rocks and scrub domes removed and the rest moved 25 m off the road (scene 1 scrub 598 to 72). Night fog halved, film lights 40000 to 26000 cd with less haze. The red carpet gets a fabric texture and a dark hem.
• 20:43Camera. Hero-frame motion blur cut from 0.35 to 0.12; three of three sampled hero stills sharp (round four: one of three). Scene 2's lens waits 18 m past the truck gap; scene 4's swings 40 degrees off the nose near fire so the ring is in frame.
• 20:43Reverted by their own builders. Moving scene 1's pyro charge (it knocked the glass down before a perfect run arrived), a camera tilt for the scene 3 jump (it cropped the car), and a darker exposure on the result orbit (the car stayed exactly as yellow).
• 20:50Integrator. No code changes needed. Four-scene autoplay 31 pass / 0 fail, every scene THEATRICAL and DIRECTOR'S CUT, $2,166,620, BLOCKBUSTER. Sixteen sloppy and lazy runs: no stuck takes, no wrecks. Nothing reverted at the gate.
• 20:51Packaged. Linux Shipping zip, 370 MB, leak check clean, Linux README inside. Packaged four-scene autoplay 32 pass / 0 fail, 9.8–10.9 ms a frame on average, worst frame 46 ms.
• 21:30Content freeze.