Agentic Streams: an agent researches a topic and streams you the briefing

Feeds decide what reaches you, how it is filtered and how it is presented. Agentic Streams turns that around. You give an agent a topic. It researches the topic on the live web, collects the real clips, screenshots and charts, writes the narration, assembles a short video with VideoDB and hands you a stream to play. No feed, no scrolling, nothing ranked by an algorithm you did not choose.

The code is open source at video-db/agentic-streams. It ships three agents, and each one is a folder you can copy to make your own.

What one agent does, step by step

Every agent in the repo follows the same five steps.

1. Research. The agent browses the web with browser-use, reading news, social posts and video sources. Every fact and every URL comes from a page it visited. Nothing is made up.
2. Gather assets. It pulls the YouTube clips, tweets, article screenshots and charts that back up what it found.
3. Write the script. It writes a narration in a broadcast voice and turns it into speech.
4. Assemble the video. VideoDB puts the clips, screenshots, narration and music on one multi-track timeline. Nothing is re-encoded by hand.
5. Deliver. The result is a stream URL, ready to play and share.

The input is a topic. The output is a video stream.

The three agents in the repo

Content creator

Research any topic and produce a one to two minute briefing. This agent also reviews its own work: after the video is assembled it indexes the output with VideoDB and checks, scene by scene, that what the narration says matches what is on screen. Where the two disagree, it fixes the cut.

News digest

Research a topic across several sources at once, YouTube, tweets and articles, and deliver a three to four minute report in a broadcast format. Ask it for Create a video report about "climate summit 2026" and it comes back with the stream.

Financial market analysis

Investigate the markets for a date, with charts, screenshots and verified clips as the evidence. Ask it for Create a market report for 2026-04-01. The length follows the amount of evidence.

More outputs from the same agents

The same three agents, pointed at other topics.

Running it yourself

You need Python 3.8 or later, a VideoDB API key (free to start at console.videodb.io) and a coding agent that can use skills.

npx skills add video-db/skills
export VIDEO_DB_API_KEY=your_key_here

Then open one of the agent folders, read its README, and give the agent a topic. It researches, gathers, writes, assembles and returns the stream on its own. There is no manual step in between.

Making an agent of your own

Each agent is one folder with three files: SKILL.md, AGENTS.md and README.md. Copy an existing folder, change the research brief and the output format, and include a working example. That is the whole recipe. The five steps above stay the same, so a new agent is mostly a new brief: what to look for, which sources count, and how long the briefing should be.

What VideoDB does in this

The agent does the reading and the writing. VideoDB does the video: it ingests the clips and screenshots, composes them with the narration and music on one timeline, serves the result as a stream, and, for the content creator, indexes the finished video so the agent can check the narration against the frames before it ships. The same primitives run the launch video Claude edited and the YouTube search agent.