Roblox turns its built-in AI assistant into an agent that plans, builds and tests games on its own

Roblox has expanded its built-in AI assistant into a working agent. It no longer just suggests code: it can break down a project, propose an action plan, generate objects, check the result and refine it based on test outcomes. On a platform where a huge share of creators arrive without deep engineering training, this is a notable shift from autocomplete toward something close to a full development co-author.

The main new feature is called Planning Mode. Instead of answering a single request with one chunk of code, the assistant first analyzes the existing codebase and the game's data model, then asks clarifying questions and turns the discussion into an editable action plan. The developer can review the plan, adjust it and only then start the implementation. In effect, Roblox is trying to turn the assistant from a snippet generator into a system that thinks about its approach to a task first. The difference matters: asking an AI to write a function is one thing, and entrusting it with designing a solution inside a live project, with its dependencies, constraints and accumulated logic, is quite another.

The second line of updates concerns content. The Procedural Models tool will let developers create 3D objects described not by a static mesh but by parameters in code. If a developer needs a bookcase, for example, they can set the number of shelves, the height, the materials and other properties through a text prompt, then change them without remodeling by hand.

Roblox is betting not simply on generating images or decorative assets but on parametric design, where an object keeps the internal logic of its shape. A staircase understands the relationship between height and steps; a table, the relationship between tabletop and legs. For indie developers and small studios, this could noticeably cut the time needed to create variations of the same object.

Separately, the company is adding Mesh Generation, which creates fully textured 3D objects directly in the game world from a text description. The feature is built on top of the Cube foundation model. In February 2026, Roblox had already demonstrated 4D generation based on Cube: objects gained not just shape but interactivity, so they behave correctly inside the game rather than remaining motionless props. According to the company, developers generated more than 160 thousand objects during early access, and projects that used 4D generation saw average playtime rise by 64 percent. It is one of the few cases where Roblox ties AI tools not only to convenience for creators but also to a measurable effect on player engagement.

The most important part of the update is the so-called self-correcting loop. The assistant can now test individual aspects of a game, find problems, propose fixes and feed the test results back into the next planning cycle. Roblox calls this agentic loops: repeating cycles of planning, execution, testing and refinement with gradually less human involvement. Next, the company wants to let several AI agents work in parallel in the cloud, not only inside a local Studio session.

Alongside this, Roblox Studio has gained a built-in MCP client through which the assistant will be able to connect to external services and tools such as Claude, Cursor and Codex. The long-term goal, which Roblox has talked about since opening the Cube model in March 2025, looks ambitious: a developer describes a game in natural language, and the system helps assemble assets, environment, code, animations and interactive behavior into a single whole.

The business logic is obvious too. Roblox operates at a scale that allows heavy investment in such tools: in the fourth quarter of 2025, the platform's daily audience reached 144 million users, against 85 million a year earlier, and the monthly audience grew from 280 to 380 million. Revenue for 2025 was 4.9 billion dollars, 36 percent more than a year earlier, and the 2026 forecast is in the range of 6–6.2 billion dollars. Total Robux purchases in 2025 reached 6.79 billion dollars.

Amid the vibe coding boom and a sharp drop in the barrier to entry for game development, Roblox is trying to solve two problems at once: give more creators the tools to launch projects, and avoid drowning in a flood of low-quality raw content. That is why the bet is not on one-off generation on request but on a more structured process with a plan, verification and iterations.

For the market, this is a signal that the next stage of AI tools is not merely assistants that insert code or draw assets on request, but systems capable of taking over a piece of the production process. If Roblox brings this approach to stable quality, solo creators and small teams will be able to assemble playable projects faster and spend less time on routine work. But the main question remains open: will the platform get truly better games, or simply more games? The answer, it seems, will become visible within the next year.