Cole Medin spent the past eight months building Archon, his open-source "harness builder for AI coding" that connects many coding agents together to work autonomously on longer tasks. It now sits at almost 23,000 stars, and he has noticed a trend: people are using it for agentic workflows that have nothing to do with coding. To show what that looks like, he builds an AI content factory that combines Archon workflows with Higgsfield, a video generation platform, to turn a still-image product catalog into a catalog full of product renders and UGC-style video ads. He is explicit that this is a demonstration, not a production-ready polished system.
Cole opens with the elevator pitch: Archon is the open-source harness builder for AI coding, letting you create larger systems where many coding agents are connected to work on longer tasks autonomously. With almost 23,000 stars, a lot of people have gotten their hands on it, and an interesting trend surfaced: they are using it for any kind of agentic workflow, including research tasks and content creation. This video is his example of Archon doing something other than coding. He flags up front that it is not production ready or extremely polished, just a look at what is possible.
The build is an AI content factory: like a dark factory, but for generating marketing content at volume, driven entirely by Archon workflows plus Higgsfield. The starting point is a common situation, a product catalog where every product is just a still image. Paying for or shooting marketing video yourself is painful, so AI does it. He shows a 10-second rendering of a product (basic, but usable as B-roll or on the product page) and a UGC-style ad with generated audio, where the product renders perfectly. Higgsfield wanted to work with him on a video, which he saw as the perfect opportunity to showcase a non-coding Archon workflow.
Creating a video with Higgsfield is as simple as sending a prompt describing what you want; more detail is better. You can do it in the platform after signing up, but they also have a CLI, which means the coding agent never has to leave the terminal. Cole keeps the two install commands in the repo README. The repo also contains a Higgsfield skill ("use this when the user wants to generate an image, video, or a UGC-style video ad") that is really just instructions for the agent on how to use the CLI effectively. He pastes a prompt asking for one 10-second vertical UGC ad of a guy in a kitchen holding the branded tumbler, using the branded product image as the reference, so logos are not generic. The agent loads the skill, calls the CLI, and returns a clip with audio that he judges "actually pretty good" for a single prompt, if a little cheesy.
Prompting once for one video is fine, but you do not want to repeat that dozens of times, and you cannot ask a single coding agent session to generate a large batch of videos, because it gets totally overwhelmed by the task. That is where Archon comes in: an orchestrator that fans out to parallel workers, each taking a single video at a time, and validates everything before handing back the finished videos. There is also an economic constraint: you want to validate the idea for the video before spending credits rendering it, so the process is image, then validate, then generate the video.
Two workflows guide the whole process. The first explores the products and generates ideas. The second, the render workflow, takes the approved ideas and creates the videos. Both use the Higgsfield CLI, and both are just Ralph loops, the same harness pattern used for AI coding, applied to content. Instead of the task list being things to do in a codebase, it is the individual products that need content. The worker prompt tells the agent to look at the next item in the queue; if nothing comes back, it outputs "complete" to exit the loop, which is exactly how a Ralph loop terminates. Cole notes he does not need to get into the YAML weeds and links a separate Archon guide for that.
All the workers share the same prompt, and five run at a time in the explore workflow. With 20 products that means four passes of five; with 13 products it runs five, five, then three. The workers do not generate video immediately, they generate an image of the idea first. The agent does its own scoring, gives recommendations, automatically scraps images that are really bad, and surfaces only the good ones. Those go into a review queue, and the ones a human approves land in the approved queue as the input to the render workflow. The render worker prompt reads: you are one worker on the render line, each run you animate one approved ad, then stop, with the Ralph loop continuing on the next video in the next agent. It claims the next approved concept via a couple of Python scripts that read the markdown documents, and prints none when there is nothing left to render. This workflow runs only three workers in parallel, deliberately fewer, because rendering actual videos costs a lot more credits, which is exactly why the image-then-video sequence exists.
Cole runs a live demo on a trimmed second version of the site with only a couple of concepts, to keep it fast. The coding agent loads the Archon skill and kicks off the workflow via the Archon CLI, which in turn uses the Higgsfield CLI. This is how Archon workflows work: your primary coding agent (Claude Code, Codex, whatever) is the orchestrator, launching the workflow and then checking in on it over time to make sure it is going well. The workflow creates the concepts, and a call to Gemini validates and scores the images. Once images exist for both products, it is his turn to approve. He approves off camera, then runs the content factory render workflow and watches it in the Archon logs, where the graph view shows the orchestrator, the workers, and progress toward finishing.
Back in the catalog, the products have gone from images only to having videos. He plays the Onyx kettle ad ("This kettle made my pour over so much better. The gooseneck gives you a perfect slow controlled pour. If you're serious about coffee, get it.") and a second, better version generated off camera, and remarks that you cannot really tell it is not a real person. He is careful to say this catalog is a demo: a real product listing would not show ads and videos inline like this, so think of the page as an admin view of everything you approved and generated, which you then carry forward into real product pages and real ads. One thing he almost forgot to show: the approval step is built into the UI, so you approve the vetted images directly on the page, which updates the local markdown document that becomes the input to the render workflow.
The takeaway he wants to leave: ideas from AI coding, harness engineering, and the Ralph loop transfer directly to other domains, here content marketing with Archon plus Higgsfield. The whole thing, Archon workflows, the catalog site, sample videos, the explore and approved queues, and the Higgsfield skill, is in a repo linked in the description and free to play with.