I am experimenting with a personal autonomous agent for YouTube production. The goal is to give the system a video idea and let it coordinate the work required to research, write, create assets, assemble the video, verify the result, and eventually publish it to the Tomorrow Without YouTube channel. This is an active experiment, not a finished hands-off product. I am building it around OpenMontage as the end-to-end production pipeline, with custom skills that define how each specialized part of the job should be performed. The production problem A finished YouTube video is the result of many connected tasks. Even when AI helps with individual steps, a creator still has to move information between research, scripting, narration, visuals, editing, quality checks, metadata, and publishing. The experiment asks a more useful question: can one agent coordinate those stages without turning the workflow into an unreviewable black box? The target pipeline is: Turn an idea into a clear video brief. Research the topic and preserve source evidence. Write the script and define the visual plan. Produce narration, imagery, footage, music, and sound effects. Assemble and render the video through OpenMontage. Inspect the final media and publishing package. Publish to YouTube only after the required checks and approval. Why OpenMontage is the production layer OpenMontage gives the workflow one place to manage projects, media, rendering, and publishing artifacts. The autonomous agent can coordinate the process, while the production system keeps the files and outputs inspectable. That separation matters. The agent should decide what work needs to happen; the production pipeline should make the result deterministic, reproducible, and testable. Every video remains a project with traceable inputs rather than a collection of temporary prompts and downloads. Custom skills as reusable production knowledge A general-purpose agent does not automatically understand the standards of a good YouTube production. Custom skills provide focused instructions for recurring jobs. The skill layer can cover areas such as: research and source evaluation script structure and pacing narration and pronunciation media selection and provenance thumbnail and title preparation video composition and rendering visual safe zones and caption placement technical and editorial quality checks YouTube metadata and publishing Each skill gives the agent a narrower contract. This is easier to test and improve than one large prompt responsible for the entire production. Creating all the supporting assets The intended workflow does not stop after generating a script. It has to resolve or create every asset required by the production plan: narration, images, video clips, music, sound effects, captions, thumbnail artwork, title, description, and publishing metadata. Assets also need provenance. The system should record where an asset came from, how it may be used, and which version reached the final render. Missing or unclear rights should block that asset rather than silently passing it into the video. Quality gates before publishing Autonomy should not mean publishing unchecked output. The pipeline needs gates that can stop the run when something is incomplete or unsafe. Useful checks include: factual claims remain connected to their research sources narration matches the approved script required scenes and assets are present captions stay within readable safe zones video and audio metadata match the target platform the encoded file plays for its full expected duration the title, thumbnail, and description accurately represent the video no upload occurs when a required approval is missing Publishing credentials should remain isolated from research and generation steps. Only the publishing stage should be able to use them. What autonomy means in this experiment The agent should be able to resume work, identify a failed stage, and retry the smallest safe unit instead of starting the entire production again. It should also produce a readable status record showing what was created, what was verified, and what still needs attention. For now, I consider human approval before upload a useful production boundary. The workflow can automate the repetitive work while still making the final publishing decision deliberate. Current status The project is being tested through the personal Tomorrow Without workflow. OpenMontage provides the production foundation, and the custom skills are being developed as individual capabilities are exercised in real videos. The important measure is not whether the agent can produce one impressive demo. It is whether the same process can create multiple videos reliably, surface failures honestly, and preserve enough evidence to understand every output. Honest limitations Research quality, generated media quality, licensing, model availability, and platform APIs can all affect a run. YouTube publishing also introduces account permissions, quota limits, processing delays, and policy requirements that the agent cannot ignore. Creative judgment remains difficult to automate. A technically valid video may still have weak pacing, an unclear story, or a misleading thumbnail. Those failures need editorial review and better skills, not a claim that autonomy has already been solved. What I am learning The hardest part is not generating individual assets. It is coordinating them with clear state, repeatable standards, and safe transitions between stages. This experiment is helping me treat an AI agent as a production operator with defined responsibilities—not as a single prompt that somehow does everything. As the custom skills mature, I will document what worked, what failed, and which decisions still require a human.