Introduction

Most studios bolt AI onto an existing engine and call it a day. Mark Kern just ripped Unreal out from under Em-8ER and replaced it with Kun-Zhi — a ground-up C and Vulkan core he claims was built with AI in six months and now moves production 93× faster. AMD and Razer are listed as partners, backer access is open again, and the pitch is blunt: stop licensing a generalist engine when your pipeline can be AI-first from the substrate up.

The Magic: Engine as AI Substrate

Kun-Zhi (Tibetan for the ground everything springs from) is not a plugin layer. Kern describes a production engine written mostly in C, with some Elixir, on Vulkan — networking, collision meshes, character animation, outfit swaps, physics-driven effects, client/server player capsules, and pilotable vehicles already in the current update.

The workflow is the story. He rotates Codex, Grok, and Claude (mostly Astra lately), but puts “vibe” in scare quotes on purpose. Production-ready output still needs someone who knows how games are structured and how to steer the models. In his words, data structures and architecture matter more; the coding was always the drudge part.

Pro Tip

Treat AI-first engines as architecture problems, not prompt contests. If your ECS, netcode contracts, and asset exporters are crisp, agents ship features. If those are mush, you get impressive demos and a maintenance tax.

What This Means for Game Pipelines

Compare this to Unity’s official Claude Code plugin or prompt-to-play platforms: those accelerate inside a known engine. Kun-Zhi argues the engine itself should be designed for agent loops — public release planned for backers first, with corporate sponsorships on the table.

Practical takeaways if you are evaluating an AI-native core:

  1. Measure feature velocity against a known Unreal or Unity baseline on the same game systems (animation export, net sync, vehicles), not vanity FPS counters.
  2. Keep a human architecture gate: agents write the bulk, seniors own invariants (determinism, rollback, memory budgets).
  3. Prefer engines that expose Blender/asset tooling as first-class AI workflows — Kun-Zhi’s AI-assisted Blender anim exporter is the kind of seam that compounds.
  4. Demand replayable builds. Hand-picked capture looks great; backers and partners will pressure-test jank.

AI-first engine checklist- Explicit world/net state contracts agents can edit- Asset exporters in the agent loop (not hand-only)- Human review on architecture, not every line- Benchmark vs Unreal/Unity on the same features

Conclusion

Kun-Zhi is a loud bet that the next leap is not prettier Unreal plugins — it is engines born for agent-driven production. Em-8ER is the living proof bed, AMD and Razer give the claim corporate gravity, and the public backer window is the stress test. Watch the velocity claims against real multiplayer play, then decide whether your studio still needs a licensed monolith.

—Titus