Codexが生んだゴブリンたち、その制作背景をOpenAIが公開 Where the goblins came from

OpenAI has published a behind-the-scenes blog post describing how a cast of "goblin" characters was created using its agentic coding tool, Codex. The piece is framed less as a product update than as a creative showcase, illustrating how a system marketed primarily for software engineering can also be applied to visual and artistic work.

According to the post, Codex appears to have been involved across the full pipeline, from initial concepting through code implementation to the generation of the final visuals. Rather than treating the AI as a one-shot image generator, the team seems to have used it as a collaborator that writes and iterates on the code that ultimately produces the goblins on screen. The result is a workflow in which character design, animation logic, and rendering all emerge from generated source code rather than from a traditional illustration pipeline.

Codex itself has evolved considerably from its origins as the model behind early versions of GitHub Copilot. Over the past year, OpenAI has repositioned Codex as an agentic, cloud-based coding environment capable of handling multiple tasks in parallel, running its own sandboxes, and completing longer-horizon engineering work with limited human supervision. The goblin project can be read as an attempt to demonstrate that those same agentic capabilities translate to creative coding, where exploration and rapid iteration matter as much as correctness.

Playful character work of this kind sits naturally at the intersection of generative art and procedural generation. Creative coding libraries such as p5.js, pixi.js, and three.js have long been used by artists to turn small pieces of logic into rich visual systems, and an LLM-based agent that can write, run, and refine such code lowers the barrier to that kind of experimentation. In this setup, the AI is not just suggesting snippets but effectively driving a sketch-and-refine loop that previously required a human programmer at the keyboard.

The broader industry context is that nearly every major coding agent is now being pitched, at least in part, to non-traditional developers. Anthropic's Claude Code, Cursor, Replit's agent features, and a growing number of smaller tools have all added capabilities aimed at designers, educators, and hobbyists who want to build interactive or visual artifacts without managing a full development environment. Positioning AI agents as creative collaborators rather than pure automation tools has become a recognizable trend, and OpenAI's goblin post appears to reinforce that narrative.

There is also a marketing dimension worth noting. A blog entry about whimsical creatures is far more shareable than a changelog, and it gives OpenAI a way to humanize a product whose core audience is professional engineers. By foregrounding the artistic output rather than the underlying prompts or scaffolding, the company can highlight Codex's range while sidestepping more contentious debates about autonomy, code quality, or the displacement of human developers.

Technical specifics, such as which exact models were invoked, how much of the code was hand-edited, and whether the generated assets or source will be released publicly, are not fully clear from the framing of the post and would need to be confirmed against the article itself. It also remains to be seen whether OpenAI intends to package this kind of creative-coding workflow into a more formal template or template gallery within Codex, or whether the goblins are simply a one-off demonstration.

Either way, the piece fits a pattern in which frontier AI labs increasingly use small, charming side projects to communicate the capabilities of their agents. For Codex specifically, the goblins suggest that OpenAI sees value in showing the tool not only shipping pull requests and fixing bugs, but also producing the kind of low-stakes, visually expressive output that has historically belonged to creative coders working alone with a canvas and a few hundred lines of JavaScript.