Google's Antigravity Agent Uploaded 347 Videos to YouTube Around the API: Up to 102 a Day Instead of 3–5
The author needed to upload 347 episodes of the "Neuro Roast" series to YouTube, but the official API allows only 3–5 videos a day, which means months of waiting. Google's Antigravity agent proposed a workaround on its own: the author logged into the account once, and the agent then clicked through the YouTube website in the background like a human, uploading up to 102 videos a day. The takeaway: if the API is too restrictive, a coding agent with browser access can be handed the entire chain.
AI-processed from TG: @cgevent (Метаверсище и ИИще); edited by Hamidun News
A familiar situation: you need to upload hundreds of files to a service, but the official API lets you do only a few per day. Usually a developer either resigns themselves to months of waiting or writes a fragile script. The author of the Telegram channel "Metaverseishche and AIishche" ran into exactly this.
He needed to move 347 episodes of the "Neuro Roast" series to YouTube, and the API, according to the agent, allows only 3–5 videos a day. Google's Antigravity agent suggested a workaround itself: it operated a logged-in browser in the background and uploaded up to 102 videos a day. This is a live example of a coding agent on a consumer subscription taking a task from data export all the way to deployment.
What had accumulated and how the agent assembled the archive
The series is a year and a half old. Over that time the channel collected a lot of posts: subscribers' work, reactions, views, and links to YouTube. The author decided to turn all of it into a proper archive and handed nearly all the work to the agent.
First, the agent was asked to go through the channel and export everything tagged "Neuro Roast" into a database. It saved reactions and views, and for the oldest posts, where the title and author were not specified, it tried to reconstruct those fields itself. Then it found posts linking to YouTube in good quality, downloaded the videos with ytdl, and stored them next to the database. According to the author, unlike GPT, it did not bombard him with clarifying questions and simply got on with the work.
How the API limit was worked around
Next, the author asked whether all of this could be uploaded to YouTube without storing the videos locally or on hosting. The agent replied that the API would allow 3–5 videos a day, and there were 347 of them. It then, on its own initiative, opened a browser window and asked him to log in. It saved the cookies and session data and continued clicking through the YouTube interface in headless mode, meaning without a visible window.
The author stresses that this is neither Computer Use nor MCP. The agent launched Chrome on his machine and pressed buttons on the pages. Empirically, they found that 102 videos a day could be uploaded to a channel with history this way. For this, the author used an older channel with about 1,500 subscribers. The channel is not yet tidied up: all episodes were simply uploaded in chronological order.
Site, deployment, and nightly automation
While the upload was running, the agent was asked in a separate chat to build an analytics site from the database. It produced the first version in about seven minutes, and it was then refined by hand. The agent came up with the dashboards itself. It noticed that the series has haters, analyzed the time series and reaction counts, and wrote an algorithm that subtracts their influence from the statistics.
To move to hosting, the agent requested SSH access, checked what was available there, rewrote the site in PHP 8.3, cut JPG previews from the videos, migrated the SQL database, and produced a link. The SSL certificate took the longest to sort out. The final site weighs 237 KB, and the videos play directly from YouTube. At the author's request, the agent also picked the frames itself and assembled a 1280×720 teaser.
Finally, a nightly task was set up. It checks the channel, finds a new "roast," downloads it, re-downloads a high-resolution version if needed, updates the database, syncs it with the hosting, and uploads the video to YouTube. All of this, the author claims, works without the API and fits within the Google AI Pro subscription at $20 per month.
What to take away in practice, and caveats
The first takeaway: an agent with access to a terminal, SSH, and a browser can cover the entire chain, from downloading to scheduling. The second: if the API is too narrow, the agent can switch to driving the interface. This is worth considering when you estimate timelines for migrations and internal archives.
There are several caveats. All the figures come from the author's own account, and there is no independent verification. Controlling a site through a browser is fragile: if the markup changes, the script breaks. Saved cookies give the agent broad access to the account, so they should be stored separately and their permissions limited. Automating the interface may violate YouTube's rules, which is better checked in advance. Finally, the author himself calls the project a toy rather than an industrial system. The agent's output must still be verified: read the logs and check what was actually uploaded.
- Nov 21, 2026A swimming team captain handed ChatGPT three scheduled tasks and cut her routine work by a multiple
- Nov 21, 2026MIRA Neural Network Renders Rocket League at 20 FPS, but Ping Hits 460 ms
- Nov 21, 2026Google built a closed copy of the internet with fake banks to train Chrome's AI agents
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
Key Questions Answered
What had accumulated and how the agent assembled the archive?
The series is a year and a half old. Over that time the channel collected a lot of posts: subscribers' work, reactions, views, and links to YouTube. The author decided to turn all of it into a proper archive and handed nearly all the work to the agent. First, the agent was asked to go through the channel and export everything tagged "Neuro Roast" into a database. It saved reactions and views, and for the oldest posts, where the title and author were not specified, it tried to reconstruct those fields itself.
How the API limit was worked around?
Next, the author asked whether all of this could be uploaded to YouTube without storing the videos locally or on hosting. The agent replied that the API would allow 3–5 videos a day, and there were 347 of them. It then, on its own initiative, opened a browser window and asked him to log in. It saved the cookies and session data and continued clicking through the YouTube interface in headless mode, meaning without a visible window. The author stresses that this is neither Computer Use nor MCP.
What are the key details on Site, deployment, and nightly automation?
While the upload was running, the agent was asked in a separate chat to build an analytics site from the database. It produced the first version in about seven minutes, and it was then refined by hand. The agent came up with the dashboards itself. It noticed that the series has haters, analyzed the time series and reaction counts, and wrote an algorithm that subtracts their influence from the statistics.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.