Zar Cabin Case
I directed a crew of AI agents to make a 33-second film and a live 3D page for an aluminium cabin case. The agents did the labour: scene and render scripts, splat training, web code and tests. The taste, the calls and the checks were mine.
Live · travel.zar.sh
Turn the case, switch finish, recolour, add stickers, share an image. It all runs in your browser.
This configurator is the version I chose to finish. Nothing went live on travel.zar.sh without my sign-off.
How this was made
Director and crew
Agents produce fast; the hard part is knowing whether the result is good. That is where I spent the time. Nothing counts as done on an agent’s own word: a captain turns my intent into briefs with pass/fail checks, crews build in parallel, a separate reviewer checks, and I sign off.
Director
Me: intent, taste, scope
Captain threads
Plan, set pass/fail checks, brief crews, check every result
Crews, in parallel
Independent review
A separate reviewer passes or fails each candidate
Sign-off
Ships only after my “ship it”
PLEASE VERIFY INDEPENDENTLY (priority order): A. Share on all 4 models x s2/s5 x 3 formats … look judged by you … floating? shadow/reflection believable? halo/cut-out edges? sticker visibility? small print present and legible? host text? … exact sizes, <=1.5 MB, QR decodes ([…]) to the same config, QR off = no QR; … Esc/focus return, camera/viewer state restored after closing (colours, stickers, orbit), no GPU device leak (max 1 live), repeated open/close x5. B. Fallbacks: … quiet notice, … studio image valid with a decoding QR, never a blank/broken image, never an uncaught page error. … Return PASS / PASS-with-advisories / BLOCKED + repro; do not fix anything.
A few of the things I caught
A handful of examples from many review notes on the agents’ renders, and what came back.
Calls I made
Each of these was a decision; the agents carried it out.
- 01
Cycles over V-Ray 31 Oct
Realism came from shading, not from the renderer: both looked equally convincing on one test scene. Cycles also runs on every machine in the render farm, so I kept it and put the effort into shading and wear.
Cycles 
V-Ray - 02
Wall B Film review
The stronger aged-plaster wall for the final room.
Plain 
A · subtle 
B · chosen - 03
A full-quality splat 7 Nov
I wanted one case you can turn all the way round. Brush trained Cycles renders of it into soft, coloured 3D blobs: a Gaussian splat. The browser sorts the blobs by depth as you turn it and blends them on the GPU (WebGPU, WebGL2 fallback). The light is baked in, so it looks rendered but stays light: the titanium case is one 6 MB file.
I kept it full quality and accepted a 6 to 7 second first load on simulated fast 4G, behind an honest loader, instead of shrinking it.
Pipeline
Agents built every step; the brief, reviews and go/no-go were mine.
- 01 →
Model
Blender. Titanium or Black, New or Worn; dents modelled.
- 02 →
Render
Cycles, split across a three-machine farm.
- 03 →
Splat
Trained into Gaussian splats with Brush.
- 04
Web
Static page, WebGPU with WebGL2 fallback, four languages.
Every release was frozen, hash-checked, reviewed and previewed privately before going live. Never from main. The old address redirects in one hop, and other sites cannot frame the page.
Render farm
I didn’t want to wait on one machine, so I designed the farm with the agents; they built it. A queue hands out chunks of frames, and Cycles runs on all three machines, so they render at once; every frame comes back to be checked before comp and grade.
What it took
A 33 s film and a live 3D page in four languages.
The live page reached version 1 in about two and a half days, through 21 frozen release candidates. The agent work that shipped came to ≈ $613 in API tokens.
I run on subscriptions; this is list price, not what I paid. With exploration and dead ends, the full total was about $778.
Hours of active agent work (about 50 in all, 31 Oct to 10 Nov). Each bar is 4 hours.
Show as table
| Active hours | Opus 5.5 | Sonnet 5.5 | GPT-6.1 Sol | Total |
|---|---|---|---|---|
| 0–4 h31 Oct | $49 | – | $2 | $51 |
| 4–8 h31 Oct | $50 | – | $4 | $54 |
| 8–12 h31 Oct to 1 Nov | $44 | – | $8 | $52 |
| 12–16 h1 Nov | $31 | – | $5 | $36 |
| 16–20 h1 Nov to 4 Nov | $39 | – | $7 | $47 |
| 20–24 h4 Nov to 7 Nov | $28 | – | – | $28 |
| 24–28 h7 Nov | $55 | – | – | $55 |
| 28–32 h7 Nov to 9 Nov | $56 | $24 | – | $80 |
| 32–36 h9 Nov | – | $37 | – | $37 |
| 36–40 h9 Nov | – | $40 | – | $40 |
| 40–44 h9 Nov | $21 | $23 | – | $43 |
| 44–48 h9 Nov to 10 Nov | $37 | $37 | – | $74 |
| 48–50.4 h10 Nov | – | $16 | – | $16 |
| Model | Role |
|---|---|
| Claude Opus 5.5 | Captains, 3D crews (animation, lookdev, splats) and release reviewers |
| Claude Sonnet 5.5 | Captains, plus web layout, copy and share-flow crews |
| GPT-6.1 Sol (Codex) | Codex crews: textures, environment and light, web prototypes |
Each session’s log was sorted by whether its output shipped; anything unclear counts as exploration. List prices of 13 Nov 2026. Harness: T3 Code, driven by my agent Zar.
Turn the case yourself.
Personal study. Not for sale. Not affiliated with any luggage brand. Worn is modelled geometry, not a photo. Finished 10 November 2026.