Autonomous Feature-Generation Pipeline for Mobile Gaming
From a one-line brief to designed, built, and tested game features
The Challenge
A top hyper-casual mobile game studio was still burning senior time on the low-leverage parts of shipping live-event features: writing design docs, generating and cleaning art, wiring UI, and testing. They wanted an autonomous pipeline that could take a one-line feature idea to buildable, tested game code, without becoming an unaccountable black box designers couldn't trust.
The Solution
We built an autonomous, end-to-end feature-generation pipeline, orchestrated by an LLM agent (Claude Agent SDK) driving a set of Model Context Protocol tool servers for the game engine, 3D DCC, design system, and generative art. From a short brief it produces a full design document with grounded research, on-brand UI mockups, generated 2D and 3D assets, engine code committed to the real game repository, and automated tests. A human approval gate sits between design and build, and each build stage self-corrects in a closed loop until it passes a five-layer verification stack: deterministic geometry checks, rendered-pixel computer vision for contrast and overlap, structured vision-model review, an automated play-mode smoke test, and real-boot integration contracts. Every run is versioned and reproducible, pinned to an exact code commit and input, so any output traces back to exactly what produced it.
Results
- Autonomous phases (with human approval gates) from brief to designed, built, and tested feature
- Five-layer verification stack; a runtime integration contract caught 5 gameplay bugs every visual check missed
- Pilot: a feature's art set generated and machine-validated in under an hour versus a ~2-hour-per-asset manual baseline
- Fully reproducible, versioned runs pinned to an exact commit and input