The Harness Behind the One-Person Game Studio

By Vatsal Bhardwaj · Published on Aug 7, 2026

Last year, when we launched Jabali Studio, we wrote that the next great game studio isn't a company - it's YOU. That post was about what we believe: creativity should belong to everyone, and imagination should be playable. The who and the why.

This one is about machinery, the what and the how.

Jabali is a Creative AI technology company, and games are our first application. Jabali is not a game engine with AI sprinkled on top.

A Model Is Not a Product

Everyone has seen the demos from each new frontier model release. A frontier model writes a snake clone in one shot and the room gasps. A great parlor trick. Then you ask for the game you actually imagined, with your characters, your art, your mechanics, your weird ideas, and it falls apart. The code doesn't run. The art doesn't match. The tenth request forgets the third. That distance between an impressive demo and a dependable creative partner is the defining gap of this era of AI.

Bigger models don't close it. Harnesses do: the context models see, the tools they hold, the sandboxes they act in, the workflows that sequence them, the caches that ground them, and above all the evaluation loops that tell you, continuously and in numbers, whether any of it is getting better. Games are a uniquely severe and gradable domain, so a harness forged there is stronger than one forged in pure text.

So for the past twenty-four months, while the industry argued about which model was smartest, we made a quieter bet: the lasting advantage would belong to whoever built the best harness for a domain that truly stress-tests AI.

Games Challenge

Playable game is the most demanding artifact you can ask an AI system to produce.

Here's what has to be true for a single game to work. The code must compile and run - an objective, unforgiving bar. The art, narrative, sound, and mechanics must be consistent with each other; a robot exploring abandoned cities shouldn't get a medieval castle tileset and a surf-rock soundtrack, unless that's the joke. The whole thing must hold together over long horizons, because a game isn't one file - it's dozens of interlocking scenes, scripts, and assets that all have to agree. And after clearing every one of those bars, the game still faces the highest bar in software: it has to be fun. There is no compiler for fun. No ground truth, no functional test.

So a game is a correctness problem, a coherence problem, a multimodal generation problem, a long-horizon agent problem - and then, on top of all that, a taste problem. Text is easy to fake and hard to judge. Games are the opposite. That combination of extreme difficulty and objective testability makes games the best training ground in software for AI harnesses. And we believe the next generation of great games will come from people who were never allowed to make them before.

Five Harnesses in Twenty-Four Months

Our conviction here isn't theoretical. We've built five generations of harnesses, and each taught us something the previous one couldn't.

Generation one was a deterministic orchestrator delegating pre-defined tasks to agents: come up with a list of characters, draft a story, generate dialogue for each scene, generate a background, configure the game UI, customize the game mechanics. A rigid process for a restrictive set of games, but reliable. The lesson? Agent orchestration was critical, but we needed more creative freedom.
Generation two was an internal copilot: an assistant inside the game engine, suggesting code, a human approving every step. It proved demand and taught us the first lesson - assistance doesn't democratize. If you have to understand the code to accept the suggestion, the gate is still up. More creative freedom, but not enough complexity or autonomy. Most industry incumbent tooling, adding AI is here.
Generation three was structured code generation: pipelines that assembled games from constrained parts, with dedicated context management deciding what each model call could see. Lesson: structure beats freestyle. Control the scaffold and models stop hallucinating architectures and start filling in craft.
Generation four was a multi-agent system - specialized agents for art, narrative, code, and design, coordinated as graphs, each owning its discipline the way departments do in a studio. This is the "living studio" we described in the Studio announcement, and it taught us that orchestration is a product surface. How agents hand off work, contest decisions, and share context matters as much as how smart any one of them is. Unconstrained creativity alongside close supervision from a creator with vision.
Generation five is where we are now. Fully autonomous agents running on our own agent SDK, operating inside isolated sandboxes, writing real code in real repositories, orchestrated by durable workflows that survive failures and retries. Nothing left to approve: the agents write the code, commit it, and hand you a playable build. A true co-producer enabling a creative vision.

Two lessons cut across all five generations. Never marry a model - the "best" model for our workloads has changed more times than we can count, sometimes overnight, and every generation of our harness got more model-agnostic. And the harness compounds while models churn. Every template, eval, failure taxonomy, and orchestration pattern survived each model transition and made the next model better for us on day one.

The Discipline That Makes It Work: Evals

The AI industry has finally converged on something game developers have known forever - you cannot improve what you cannot measure, and in creative domains you must first invent the measurement. For example, game programmers measure framerates across a standard set of configurations to understand if the game is ready to ship. Similarly, game designers will measure level drop-offs and pair it with player profiles to deeply uncover design problems.

The emerging discipline of AI evals - error analysis before metrics, failure taxonomies before dashboards, LLM-as-judge with real calibration - is the closest thing this industry has to an engineering method.

We trace everything: every creation run produces a full record of prompts, tool calls, intermediate artifacts, costs, and the final build.

We read before we count. New failure modes are found by humans doing error analysis on real traces, not by staring at dashboards. Only when a failure mode has a name does it become a metric.

We judge in layers. The bottom layers are objective - does it build, does it boot, does it run clean. The middle layers check coherence: do the assets, narrative, and mechanics agree with each other, and with the prompt. The top layer is the hard one. Is it fun-shaped - a goal, a challenge, a reason to keep playing? For the subjective layers we use LLM judges, and we treat the judges themselves as systems to validate, calibrated against human annotations, so a verdict means something beyond vibes.

We make agents play the games. Our QA harness drives generated games in a real browser - playing them, probing them, trying to break them - and LLM-based QA interrogates what happens. Games are testable in a way essays never will be. We exploit that mercilessly.

And we gate change on evidence. A new model doesn't get promoted because it's exciting. It runs the suites, per template and per modality, and the numbers decide. This is how we've absorbed wave after wave of model churn without the product regressing: the eval harness turns industry chaos into a routine engineering decision.

Vibes don't scale. Judgment does, if you build the machinery to apply it consistently. That machinery is the real technology behind Jabali.

AI First, Games First

Look at what we've had to build: an agent SDK for long-horizon autonomous work, durable multi-agent orchestration, a multimodal model gateway with routing and cost metering, semantic retrieval infrastructure, a template system for injecting domain knowledge into generation, and an evaluation harness that can score outputs no compiler could ever check. None of that is a "game feature." All of it is general AI infrastructure - what you need whenever you want AI systems to reliably produce complex, multi-part, quality-gated creative work.

The studio still stands. Now you know what it stands for and stand on.

An Invitation

If you're a creator, everything above exists so you never have to think about it. Type what you imagine; we'll sweat the harness.

If you're an engineer or researcher who got excited reading about agent orchestration, eval design, or making "fun" measurable - these are the problems we work on every day, and we're hiring people who want the hardest version of them.

And if you're building in AI and this framing resonates, we'd love to compare notes. The playbook we're writing for games is a playbook for AI products at large.

The next great game studio isn't a company. It's you - standing on the best AI harness we know how to build.