Procedural Press #003
AI + Game Dev, Weekly — Coverage window: Mar 3 through Mar 9, 2026
Scene Kernels, Steam Funnels, and Agent Ops
Structured Worldbuilding Is Still Beating One-Shot Generation
The week opened with a blunt question: are splats, video models, and diffusion-heavy worldbuilding actually converging on shippable game spaces, or just on prettier demos? The strongest answer was less anti-AI than anti-opacity. Builders were not disputing that visual quality is climbing. They were arguing that state, authorship, and gameplay semantics still live outside the generated output.
That is why the more credible workflows kept moving back toward greyboxing, editor kernels, and scene formats that preserve intent. By scene kernel, builders mostly mean the core representation the game or editor is built around: the data structure that decides how rooms, surfaces, triggers, nav data, spawn points, and other gameplay-relevant objects are stored and edited. Connected planes, brushes, spawn markers, nav paths, triggers, and editable semantic layers came up more often than "generate the whole level." Even when teams liked splats or 3D generation, they tended to frame them as pre-vis, mesh acceleration, or finishing layers around a deterministic scaffold.
The important shift is architectural. Teams are increasingly treating controllability as the core product and generation as an attached capability, not the other way around.
Practical takeaways
- Keep world structure in an explicit scene model instead of burying it in generated pixels.
- Use generative systems for reference, blockout acceleration, or asset finishing until they can preserve reliable game state.
- Pick your scene kernel early: the base scene data model tends to lock in what your tools, exports, and agent workflows can do later.
Resources
Web-First Game Stacks Are Becoming Opinionated
A second strong pattern was the emergence of a more recognizable web-native stack. Browser distribution still matters for zero-friction testing, but the technical discussion moved beyond "just ship it on the web." The real work now is in keeping game logic portable across browser, desktop shell, and self-hosted infrastructure without rewriting the project every time packaging changes.
That showed up in several places at once: browser games iterating publicly, debates over Electron versus Tauri versus Electrobun, self-hosted deployment with agent-readable infrastructure surfaces, and project teams reworking expensive dependencies like map APIs instead of treating them as permanent foundations. Even packaging questions were framed as systems design questions: where should rendering live, what can be offloaded, and what has to stay portable if the project survives its first version?
The result is a more mature view of distribution. Deployment target, shell choice, and hosting surface are becoming first-class engineering decisions instead of end-of-project chores.
Practical takeaways
- Separate core game logic from packaging and platform adapters as early as possible.
- Assume browser and desktop targets may both matter, even if one leads at launch.
- Prefer infrastructure surfaces that are observable and scriptable; they fit better into agent-assisted workflows later.
Resources
Agentic Tooling Is Getting More Operational
The most useful model conversations this week were not about raw benchmark supremacy. They were about operational leverage: generating structured issue flows, reviewing PRs with multiple models, importing large quest graphs, reverse engineering old binaries, compiling odd targets, and tightening deployment loops after release. The signal was not "the model can make a game by itself." The signal was that teams are getting better at turning model output into bounded production work.
That change is especially visible around deployment and maintenance. Builders described agent loops that read setup guides, push builds, watch player feedback, patch bugs, redeploy, and report back. Others were using one model to generate and another to review. Realtime backend tooling kept surfacing in the same conversations, which suggests multiplayer experimentation is getting easier to stand up fast enough that review and ops quality now matter as much as raw generation speed.
The winning pattern is less autonomous magic than explicit workflow design: generation, review, deployment, and feedback handled as separate stages with different tools and trust levels.
Practical takeaways
- Split generation, review, and deployment into separate steps instead of asking one model to do everything.
- Require issues, milestones, or PR artifacts so the model's work has durable structure.
- Give agents least-privilege infrastructure access wherever possible; observability is usually more valuable than full control.
Resources
Projects Worth Watching
- Hovertag: A live browser graffiti battler that is already iterating in public, with bots and timed rounds turning a neat hook into a clearer multiplayer loop. Play
- VibeSail: One of the clearest examples of a real web game forcing architectural decisions in public, as map-cost pressure pushes the stack toward caching, replacement, and self-hosted infrastructure. Site
- Zombies Per Minute: A Factorio-like factory-defense game with a playable web proof-of-concept and a sharper Steam pitch than most early projects. From the public site and Steam funnel, it appears to be using the browser build to validate the loop now while building wishlists for a planned full release. Web Steam
- Warchesst: A public tactics prototype that earned attention the old-fashioned way: by being immediately playable and concrete enough to judge on feel instead of pitch. Play
- Mac-native "Minecraft but Infinite Craft" prototype: An OpenClaw-built experiment that stood out because it crossed the line from concept demo to runnable native software on macOS. Post
- AI-native Blender prototype: A fast interface experiment that reframed the question from "can AI generate assets?" to "what should an AI-first 3D tool actually feel like?" Post
Tools & Drops
- Coolify: Self-hosted deployment that is increasingly being treated as an agent-readable control plane, not just a hobby PaaS. Site
- Electrobun: An increasingly credible packaging layer for teams that want TypeScript ergonomics without defaulting to Electron overhead. Site
- SpacetimeDB: Still gaining mindshare for realtime multiplayer backends, especially in Rust-oriented stacks. Site
- PartyKit: Realtime infrastructure keeps surfacing as model gains make small multiplayer builds easier to stand up. Site
- Qwen 3.5: Open-weight planning and coding complement for teams trying to reduce dependence on always-on hosted models. Model
- OpenClaw: Maturing from curiosity into a practical surface for review, deployment, and post-launch maintenance flows. Docs
Cost / Meta
Cost pressure changed technical decisions more than leaderboard talk did this week. Expensive map APIs, token-heavy long-context sessions, and packaging overhead all pushed builders toward replaceable components: browser-local inference where possible, shell-agnostic deployment, selective use of frontier models for review or hard tasks, and infrastructure that can be inspected without handing a model the keys to everything.
Industry News
- GDC's 2026 State of the Game Industry report landed: more than 2,300 respondents, continued layoff fallout, Unreal ahead of Unity in primary-engine share, and markedly worse sentiment toward generative AI than last year. GDC report
- Unity scheduled a dedicated Unity AI Beta day for Thursday, March 12, 2026 at GDC: alongside its core engine and IAP talks, Unity is now putting AI workflow demos directly into its main GDC programming. Unity at GDC 2026
- NVIDIA framed GDC 2026 around RTX neural rendering and AI-assisted development: its March 10 programming leans heavily on neural graphics, path tracing, and AI-enabled workflow messaging for game teams. NVIDIA at GDC 2026