Dream-loop is a Codex skill that first generates a target image, then builds toward it and uses a critic agent to check the result, and has 1,569 GitHub stars in 20 days
Dream-loop is a new open-source agent skill that tells a coding agent to first generate the target picture, then build toward it and use a second agent as critic, and has 1,569 GitHub stars in its first 20 days.
Image: achimala (GitHub)
Why it mattersA written brief leaves the agent to guess what the finished thing should look like; a target image plus a critic loop replaces that guess with a real comparison the building agent can act on.
A written brief leaves a coding agent to imagine what the finished screen should look like, and different runs imagine different pictures. Dream-loop is a new open-source agent skill that tells the agent to draw the target picture first, build toward that picture, and pass the live screenshot to a second agent whose only job is to compare the two.
The skill lives at achimala/dream-loop on GitHub, published under the MIT licence by the author who goes by @anshuc. The repository was created on 2026-09-07 and has 1,569 stars and 166 forks in its first 20 days.
The closed loop, in five steps
The README lays out the loop in five numbered steps. The AI first uses image generation to produce a high-quality target screenshot of the finished thing. It then writes and runs the code with that target in mind. A separate AI critic looks at the live screenshot beside the target and returns feedback. The building AI loops back to step 2 until the critic is satisfied. As an optional fifth step, the AI dreams up an even better target based on the current state and starts the loop again.
The README uses the wording "a separate AI critic" for step 3 and lists subagents as "optional but strongly preferred" in the prerequisites. The critic runs as its own agent so it can look at the live screenshot with no memory of the code that produced it, and only compare the picture with the target.
What it needs from the host agent
The prerequisites list three things: image generation, either built into the host agent (Codex and Grok are named) or reached through an API such as Gemini; vision input, so the critic can see the live screenshot; and subagents. Blender is optional and only needed for custom 3D modelling, and the author says the Blender MCP or scripting interface produces better results than driving Blender through a general computer-use tool.
The author states the skill has only been tested with GPT-6 Astra in Codex and says other strong models such as Claude Fable 5.1 can likely work too. Teams that pick a different model are running the skill in a setup the author has not tested.
The demo shows what the loop produces
The README includes one worked prompt asking the agent to build a browser Three.js graphics demo: isometric camera, voxel-ish art with realistic shading, reflective wet floors, a character in an interesting fantasy scene, click-to-move controls, a one-hour time limit. GPT-6 Astra in Codex on high effort produced the demo the author links, hosted at dream-loop-demo.anshu.dev, and the same page loaded a working 200 response when this piece was written.
The specific claim behind the traction is small. An agent left with a written brief guesses what a good result looks like, and different runs produce different guesses. A target image and a second agent that compares the live output to that target replace the guess with feedback the building agent can act on, one iteration at a time.
Source
- Repository and README: achimala/dream-loop on GitHub
- Live demo linked from the README: dream-loop-demo.anshu.dev
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.