Browser Use Ships game-mode, an AI Agent That Builds and Playtests Web Games

Browser Use just shipped something that feels like a party trick until you realize it isn't: an AI agent that builds fully playable web games from a text description. The feature, called /game-mode, is live on their cloud platform and is temporarily free to try while the team collects feedback.

What it actually does

The agent takes a natural-language prompt, writes the game code, deploys it to the web, and then , critically , plays it. That last part is what separates this from a code-generation tool. Browser Use's agent can open the game in a browser, interact with it like a real user, spot bugs, and iterate. As the team put it in the thread: "When agents can test the websites they create and iterate on them, the games get 100 times better."

The feature supports:

• 2D and 3D games (rendered in the browser via WebGL)
• Online multiplayer
• Two model backends: Kimi K3 or Claude Fable 5

Users in the launch thread shared working games ranging from polished 2D platformers to 3D worlds with distinct art styles. The team was handing out $50 in Browser Use credits to the first 20 people who shared a game with feedback , a clever way to stress-test the feature under real conditions.

The models powering it

The choice of two backends is deliberate, and both are genuinely frontier-tier right now. Kimi K3, released by Moonshot AI, is a 2.8-trillion-parameter open-weight multimodal reasoning model with a 1-million-token context window and always-on "thinking mode." Within hours of release, it jumped from #18 to #1 on the Frontend Code Arena leaderboard with a score of 1679, overtaking Claude Fable 5 , the model most developers considered untouchable for front-end work. That makes it a natural fit for generating browser game code.

The other option, Claude Fable 5, is Anthropic's most capable widely released model. It was built for the most demanding reasoning and long-horizon agentic work. It handles complex, multi-stage work with minimal oversight, letting teams hand off large projects and review completed work rather than supervising every step. Both models have the context window and code-generation chops to hold an entire game's state in memory across multiple iterations.

One caveat the team flagged in the thread: Kimi K3 is hitting capacity limits. They tried to purchase more API access and couldn't , a sign that demand for the model is outpacing supply right now.

Why this is more than a demo

The real breakthrough here isn't "AI writes game code." Tools like Replit Agent and various vibe-coding platforms have done that for a while. The difference is the closed feedback loop: the agent writes code, then uses Browser Use's own browser automation to actually run and play the game, observe what breaks, and fix it , all without human intervention.

This is the same architecture Browser Use uses for web automation tasks (filling forms, navigating sites, doing research), just pointed at a creative output instead of a transactional one. Browser Use is the default Python answer for browser automation , install it, hand it an LLM key, point it at a Chromium endpoint, and it reads the DOM and plans. It supports multiple providers via LiteLLM, multi-tab, and hackable internals.

The multiplayer support is particularly notable. Generating a working multiplayer game requires the agent to spin up a backend with WebSocket support, coordinate state between clients, and verify that two browser sessions actually sync , all autonomously. That's a non-trivial agentic workflow.

The broader context

This launch sits at the intersection of two fast-moving trends. First, frontier open-weight models are now genuinely competitive with proprietary ones for code generation. Kimi K3 lands fourth among all frontier models in independent testing , trailing only Claude Fable 5 and GPT-5.6 Sol, and edging past Claude Opus 4.8. Similar to how DeepSeek's R1 generated attention for being more cost-efficient relative to proprietary technologies, the interest in Kimi K3 can be attributed to rising concerns about AI's overall cost and ability to generate returns on investment.

Second, agentic loops that include self-testing are becoming the standard pattern for high-quality AI-generated software. Browser Use's own research handed Claude Code a CLI to their eval platform and let it loop for 20 cycles, tree-searching the prompt space , that's how they got 97% on Online-Mind2Web. The game-building agent applies the same principle: iterate until the output actually works.

How to use it

Access is through the Browser Use cloud platform. The /game-mode route is currently free while the team collects feedback. To get started:

1. Go to the Browser Use cloud app
2. Start a task with the /game-mode command
3. Describe the game you want , genre, mechanics, art style, multiplayer or not
4. Choose your model backend: Kimi K3 (faster, cheaper, currently capacity-constrained) or Claude Fable 5 (more reliable for complex multi-file projects)
5. The agent builds, deploys, and tests the game, then hands you a live URL

The practical use cases go beyond fun. Rapid game prototyping for jam submissions, interactive demos for pitches, educational mini-games, and multiplayer experiences for internal events are all within reach. For developers, it's also a concrete way to see how a self-testing agentic loop performs on a creative coding task , which is increasingly the architecture pattern worth understanding.