DevQA and the qa harness
Scripted play sessions with a dedicated server and several clients: screenshots, video, contact sheets and a visual diff against an approved baseline. It runs on every push.
Behind the scenes
How Wyrdale was made
Four days, one director and a crew of AI agents, with every stage checked by an independent verifier.
Wyrdale was made by one person directing a crew of AI agents, in four days, for the 2026 Epic AI MegaJam. The agents write specs, code and tests; every design call belongs to the director.
The art is AI-generated and human-directed: every image on this page started from a style brief (gouache, slate and codex), was reviewed, and was redone when it needed to be. The interfaces below are style studies, not game screenshots.
A lot of what's on this page doesn't exist yet. We'd rather show the plan honestly than promise what isn't ready.
- Coding agents wrote the C++ server and client, the Node.js platform (accounts, persistence, rankings) and the tools.
- Astra, the art director, writes the art direction, generates every reference sheet and reviews its own batches. References become 3D through image-to-3D tools.
- Independent verifiers rebuild every stage from a clean clone, run scripted play sessions recorded on video, compare screenshots against a baseline and try to break things.
- A lore keeper checks every line of world text against the world bible.
- An economy simulator runs the valley's economy before any number goes into the game.
- The AI arena plays thousands of bot fights to find exploits and balance combat.
The pipeline
How an idea becomes part of the valley
Two assembly lines ran side by side for four days, one for code and one for art. Nothing reached the public server without passing every station.
Code
- SpecThe director writes the rules in the design docs.
- AgentsCoding agents build it in C++, TypeScript and Python.
- VerifierAn independent agent rebuilds it from a clean copy and tries to break it.
- CI and QAThe build server compiles, runs the tests and plays a recorded QA scenario.
- LiveOne command deploys the server, the API and the client.
Art
- ReferenceAstra paints a reference sheet in the game's style.
- Image to 3DRodin, Tripo or Meshy turn the sheet into a model.
- CleanupBlender scripts fix scale, pivot, materials and level of detail.
- ImportGLB through Unreal's Interchange, with collision and LODs.
- In the gameA QA screenshot checks it from the game's own camera.
Made for the game
What we built to build the game
Most of the work was not the game itself but the tools that let a crew of agents build it safely and quickly.
The AI arena
Bots with the same inputs as a player fight in parallel, faster than real time: about 1,900 fights in 110 seconds, with win rates, Elo and exploit alerts.
endless-data
One source of truth for items, recipes, skills and lore. It generates C++ and TypeScript and is checked by the CI and at login.
The economy simulator
Runs the valley's economy for days and weeks before a price or a recipe goes into the game.
The lore validator
Checks every line of world text against the bible, a blocklist of names from other games, and whether each fragment can be learned in four days.
World and kit generators
Scripts that lay out the 1 km² valley, place about 37,000 pieces of nature and build the house, dungeon and furniture kits in Blender.
The animation pipeline
149 clips retargeted to one skeleton, with the locomotion graph generated by code instead of by hand.
The platform
Accounts, persistence, trade, rankings, voting, an admin dashboard and the NPC dialogue budget, in Node.js and PostgreSQL.
One-command deploy
Builds the Linux server in Epic's container and ships the API, the game server and the client zip.
The metrics script
Counts commits, assets and agent work for this site's dev diary.
The workshop
Tools we used
Engine and content
- Unreal Engine 5.8the game, driven by agents through Epic's official MCP and editor scripts
- Blender 5.2kits, cleanup and renders, through scripts and the Blender Lab MCP
- Niagaravisual effects, wired by an editor utility in C++
- Mixamo and Epic's Game Animation Sampleanimation clips and locomotion
- Poly Haven and ambientCGCC0 textures for terrain, rock and buildings
AI models
- Claude Code (Claude)the coding crew, subagents and verifiers
- Codex (gpt-6-astra)Astra, the art director, and its image generation
- ClaudeNPC conversation, with a daily budget
- Nano Bananaquick reference edits
Image to 3D
- Rodin, Tripo and Meshyprops and statues, organic shapes, characters and animals · See the comparison
Infrastructure
- Gitea with LFS and Actionsrepositories, large files and the build runner
- Docker and Epic's dev-slim imagethe Linux server build, without compiling the engine
- Fly.iogame server, API and database in São Paulo
- Cloudflare Pagesthis site
- ffmpegvideo of every QA run
From the workshop floor
The evidence behind the choices
Every important decision was measured first. Three examples, with the real numbers.
Which image-to-3D service for what
The same 7 test assets (a brazier, a crate and barrel, an anvil on a stump, a market cart, a founder's statue, a bench and a builder) went through 6 services with the same reference image. Each result was scored from 1 to 5 and timed for how long it would take to fix.
Left to right, in each strip:
- Meshy 6
- Tripo H3.1
- Rodin Gen-2.5 (API)
- Rodin Gen-2.5 (site)
- Hunyuan3D 3.1
- Trellis 2
| Service | Average score | Fix-up time | Used for |
|---|---|---|---|
| Rodin Gen-2.5 High | 4.9 / 5 | ~15 min | hard props, statues and characters |
| Tripo H3.1 | 4.1 / 5 | ~20 min | organic and colorful pieces: trees, bushes, awnings |
| Meshy 6 | 3.9 / 5 | ~22 min | first characters and animals |
| Hunyuan3D 3.1 | 3.4 / 5 | ~28 min | not used |
| Trellis 2 | 2.7 / 5 | ~38 min | not used |
Also found: the classic FBX importer broke every generated mesh, while the same files in GLB through Interchange came in clean.
Why we built the AI arena
An MMO lives or dies by its fights: enemies have to be interesting and challenging at the same time, and no way of playing can be the only right one. Balancing that by hand, a few test fights at a time, takes weeks and still misses things.
So we built an arena where bots play with exactly the inputs a player has, reaction delays included, and fight thousands of matches faster than real time. The same brains then became the enemies in the valley.
- It shaped the enemy AI: wolves that run when hurt, guards that chase criminals, a lord that gets faster below a third of his health.
- It balanced the skills: archers were too strong and mages too weak, and the arena showed it before any player did.
- It caught exploits: fights where nobody could win because both sides kept running away, and NPCs you could kill for free by walking backwards.
What the economy simulator checks
35 scenarios of the valley's economy, each simulated for 90 days with 5 to 10 random seeds, before any price went into the game.
- Buying from NPCs to resell or melt down never turned a profit, so the village reserve can't be farmed.
- Inequality stayed moderate in every scenario (Gini around 0.5).
- It also shows what still needs tuning: item prices and a newcomer's income still miss our targets in most scenarios, so the economy keeps being adjusted after the jam.
Under the hood
The platform behind the valley
The game server is only half of Wyrdale. A Node.js platform with PostgreSQL keeps everything that has to survive a restart, and gives the director a control room.
Accounts and sessions
Sign-up, login and password reset, plus the tickets that let a client join the game server.
Persistence
Characters, backpacks, banks, dungeon chests and the state of the world survive restarts. Every change goes through atomic operations.
Item audit
Every item has a unique id, so the admin can trace where it came from and who held it.
Admin dashboard
Players and characters, bans, suspensions and roles, servers and live events, in one place.
Rankings
Skills and PK counters recalculated every 15 minutes, with a public API. The site's rankings page is resting for now.
Votes, quests and the Journal
Community votes with public results, the state of every building site, quest progress and trust with each NPC.
NPC dialogue budget
Every conversation with Claude goes through the platform, under a real daily cap, so a busy day never runs away with the bill.
Jam build
What's in this cut
- A 1 km² valley: the town on stone terraces, villages, forest, river and lake, a mine and four dungeons
- Live dedicated server, accounts and characters
- Woodcutting, mining and fishing; forge, anvils, carpentry and campfire; three qualities and bandages
- Melee, bows and spells; death with full loot, ghost and resurrection
- Bank, NPC shops with opening hours, player trade, local chat, PK title and rankings
- Villagers with daily routines: they drink, dance, work the anvil and gossip at the tavern
- Conversations with the locals in your own words, through Claude, with trust and lies
- Quests and the first community project: the Arcane Tower building site
- 175 NPCs and creatures, balanced in an AI arena that has already closed real exploits
- Town votes: after the Arcane Tower, players choose the next building by bringing materials to its site
After the jam
The extended cut
- The full Journal investigation board
- Player homes: one per account, built anywhere in the buildable area with no fixed plots, with room for about 50 at first
- The arena learning from real player fights and reported exploits