Carny Labs

Some of the world's best coders say they have written a line of it for ages. Not since vibe coding came along. And vibing has got a lot more refined. I don't 'vibe' code. I call it 'clash coding'. I know enough I am one artist trying to build Carny Island — the game I have carried around for years — with two machines doing the work that would normally require a studio.

That sounds simple until you try to deploy a real agent.

The breakthrough was realising that to deploy a real agent, you have to write the job description, build the skills and tools it needs, point it to a Manual — a Bible — and then test and test it on the job. This is like interviewing humans for a real position, then training them.

I had to research all the key positions involved in building a video game. I used Claude for this research. The Level 1 phase involved writing up the skills and building tool sets for 14 roles. Easy. Huh? Nah.

The first re-build with my new team was a total, absolute disaster. The train didn't just come off the tracks. It ended up in another country. The sun came up through the basement (literally) and the art went to the rubbish dump. AI had researched all the roles but I didn't know how each role related to Carny. I hadn't given these bots any tools. I'd been a lazy human and thought AI would know what it was doing.

I had to repeat the hiring process all over again; this involved understanding each role and the work that needed to be done. The skills needed and the tools needed for each role. Each agent was created individually. It took all day and into the night for one of the world's smartest platforms to 'scaffold' each job and another 24 hours for Sol at 5.5 Ultra to review, tweak and re-build as 'super agents'.

They are now being tested on the coal face. And they are failing and we are learning from the mistakes.

This is where the difference between an impressive demonstration and a working production team becomes obvious. A job title is not a job. A prompt is not a skill. A collection of instructions is not experience.

The agent has to know what it owns, what it is forbidden to touch, which Bible is authoritative, what evidence counts and when it must stop and ask for help. Then it has to survive the work.

Yesterday the work was the Crookwood basement. The team ran a combined test with 27 green checks. The keyboard worked. Sneak and Duck worked. The camera kept its measured distance from the Villain. On paper, the candidate looked healthy.

Then the picture failed. At the north wall, the camera moved sideways to avoid the wall. The Villain is a flat, hand-drawn theatrical puppet. The sideways camera presented The Villain edge-on, and it disappeared.

The measurements passed. The picture was wrong. The candidate was rejected and quarantined. It never reached me as a supposed success.

That is the coal face. It is also the point of the whole exercise. The team is not being trained to produce green reports. It is being trained to protect the game.

This morning the clean rebuild found another piece of buried ballast: 115 old painted PNG copies and 1,938 links to them sitting underneath a world that is supposed to remain true to my vector art. The new foundation removes that fallback instead of merely switching it off. Crookwood will prepare the vector renders it needs, reuse them while the player moves, and reserve extra detail for Magnify.

Again: a renderer is not just somebody who makes an image appear. The job is to understand why the crooked lines matter, what must remain SVG, what can be disposable, and how to prove the art has not been quietly replaced.

One artist created the world. Two machines made it possible to build. But only if the machines can be hired, trained and tested like a real team. That is what we are building now.

The 14 roles continue their trials on real Carny Island work. The immediate test is the basement rendering foundation: one complete stable session, no corruption, no memory surge, no soft painted fallback and no claim of success until the moving picture survives.

Carny Island will test the waters first as a 16:9 PC soft launch on Steam and itch.io. Then comes iPad. Then the mobile version, with phones as playing pieces, and the tabletop game for Halloween. That release order is now in the Game Plan.