How the studio works

Móberg Studios is me and a team of AI agents. I decide what gets built and when it is good enough. The agents write, test and review the work under rules that are written down in the agent-built projects and enforced by tests wherever a rule can be checked. This page says what those rules are, and where they have not held.

The loop

The studio loopSix steps in order: founder intent, head-of-dev agent, specialist agents, test and CI gates, review, founder sign-off. A return arrow leads from review back to the specialist agents.1. Founder intent2. Head-of-dev agent3. Specialist agents4. Test and CI gates5. Review6. Founder sign-offrework
  1. Founder intent. I say what I want and why, in a brief, a design decision or an answer in the project’s decision queue.
  2. Head-of-dev agent. In Hollowdeep and Recon-It one agent owns architecture, the build and release pipeline, and security, and cuts my intent into small slices for the specialists.
  3. Specialist agents. Hollowdeep’s roster gives each specialist an area, such as gameplay, UI, audio, QA or art direction. In Recon-It each agent edits only the files it owns, and anything else becomes a failing test or a note to the owner. In Make Tower, Win War the agents that write code work in their own worktree.
  4. Test and CI gates. Make Tower, Win War, Hollowdeep and Recon-It verify a slice with tests before it is committed. ylfamarin.is has 18 test files that run locally, but no CI, and its deploy script does not run them. I can cite no automated tests for waldorff.is.
  5. Review. An agent with no stake in the work audits the result. What it finds is fixed before the work moves on; in Recon-It each finding becomes a failing check first.
  6. Founder sign-off. I play the build or read the result, and accept it or send it back.

Who decides what

The founder

I own the vision, the taste and the product: what is built, how it should feel, what ships and what the rules are. I play the builds and answer the decision queue. In Make Tower, Win War my standing rule is that I decide names, money, accounts and anything outward-facing.

In Hollowdeep I moved the camera from 3D to 2D pixel art, ruled out ads, timers and paywalled quality of life, and set the world structure after my own first clear. In Recon-It I set the rule that the engine proposes and a person decides.

The agents

The agents implement, test and review. Hollowdeep has eight specialist agents and a lead, Recon-It twelve roles, and ylfamarin.is seven defined agents. In Recon-It 289 of 305 commits carry a Claude co-author line, in Make Tower, Win War 267 of 361, and in ylfamarin.is all 39.

The head-of-development role holds architecture, release and security decisions, so in Recon-It records 0001, 0005 and 0009 are its decisions and not mine. Product, brand and what ships stay with me.

How the work stays honest

Small test-first slices
In Hollowdeep and Recon-It a slice is one commit, and the commit waits for a green suite. Recon-It’s rule is a failing check first, then the smallest change, then the whole suite; its settlement tests were written before the code. Hollowdeep’s suite was last recorded at 1,903 tests, run in a separate clone one Unity run at a time. The limit: Hollowdeep’s GitHub workflow has never passed, and it cannot until the Unity licence secrets are added, so for that game the green is reported by an agent on a local machine.
Architecture decision records
Recon-It keeps 11 numbered records. Record 0010, for example, explains why sign-in has no hosted provider: it would see every address, add an origin to the security policy and add a bill before there is revenue. The other projects keep a decision log in a handover file. Where a rule can be a test I prefer that to a document: Hollowdeep’s architecture tests fail the build if the simulation core reads the clock or uses unseeded random.
Independent review rounds
Reviews are done by agents that did not write the work. In Make Tower, Win War the project rules reserve the strongest model for milestones and give it the code and not the lead’s conclusions. Hollowdeep had three review agents on code, docs and QA; one found that autosave could delete a finished run before its result was banked, now fixed. Make Tower, Win War had eight role reviews, two independent-tester reviews and a read-only head-of-development review; the lead checked every finding against the code and one was wrong and withdrawn. These are agent reviews. They are cheaper than a human’s and they do not replace one.
Founder decision queues
Hollowdeep and Recon-It each keep a founder-queue file, where anything that needs me goes, in Recon-It with a recommended default. Recon-It’s studio rules state that no agent’s message is my approval. In Make Tower, Win War the rules say agents ask before deploying the live relay, and that downloads, purchases and sign-ins I approve are carried out by the lead session, because an agent cannot take my permission from another agent’s message. For ylfamarin.is the project rules say no deploy, DNS or domain change happens without my explicit approval.
Key and secret hygiene
Recon-It’s rules say agents never handle API keys, passwords, payment details or real client data, and it runs on generated data only. The keys for this site’s art tools live outside the repository and are read at run time. Agents are given the path of a key file, never its value, the code redacts keys from errors, and a test fails if a key shows up in a recorded failure.
Provenance and evidence ledgers
Every generated image on this site is logged with its tool, model, seed, size and cost, and a hard spend cap is enforced in code rather than remembered. Every number in the case studies has an entry in an evidence ledger with the command that produced it. A script re-runs the commands that are safe to repeat and fails if an output has moved; entries it cannot re-run, such as a measurement of a live site, are listed as skipped and not hidden.
Model choice by evidence
Make Tower, Win War has written rules for this: mechanical work such as QA runs, counting and format conversion goes to the cheaper tier. Agents that write code stay on the stronger one, because a mistake there costs more than the saving. Each report from a cheaper tier is logged with how much rework it needed: two clean reports keep a role there, one large rework moves it back up.

Across the projects

  • 1,048Commits across four projects
  • Over 4,000Automated tests and checks, four projects
  • 25Days from first commit to latest, four projects
  • 11Architecture decision records (Recon-It)

Each total is summed from the case-study ledgers and checked by a script. Recon-It is the only project with numbered decision records.

Identity

Móberg is the Icelandic name for the palagonite tuff that forms when an eruption meets glacier ice, and the M in the mark is two table mountains, one of ice and one of fire.

The mark builds itself: the glacier-blue peak rises, then the ochre peak, then the name Móberg Studios appears beneath.

Brand art

The logo is vector artwork built in code with the studio's AI design agent. No image generator was used, and I chose it from four rounds of drafts. The rest of the brand art on this site, which is the hero backdrop, textures, mascot, pixel icons and 404 scene, was generated with FLUX (Black Forest Labs) and PixelLab, and I picked what is used. The social image combines the logo with a generated backdrop and mascot. The generated art is decoration. Every project visual in the case studies is a real capture, and where a project’s own art is AI-made its page says so.