NotebookLM Gets an Agent OS Upgrade—Here's What

There's a particular species of YouTube video that opens with a declaration so large it could only be true or embarrassing: "This changes everything." Julian Goldie's recent NotebookLM walkthrough leads with exactly that energy. The title promises NotebookLM has "changed AI forever." The first sentence informs you that "most people are using it the wrong way."

You could dismiss this as standard YouTube inflation—and some of it is. But spend eighteen minutes with the actual workflow Goldie describes, and there's something technically substantive underneath the hype. Whether it's right for you is a different question.

The Actual Problem Being Solved

Here's the thing Goldie gets right before anything else: NotebookLM, in its default state, is genuinely a dead end for a lot of people.

You feed it sources. It synthesizes. It generates a podcast, a report, maybe a video overview. And then that artifact just... sits there. You don't know where your notebooks are relative to each other. You can't easily pipe the output into another tool. You can't edit the generated content in any meaningful way. Every time you return, you're starting fresh context.

"The old way was like you would use Notebook LM alone," Goldie says. "You'd open up Notebook LM, you might generate an audio preview, you play it once, you forget it. You don't know where all your files live."

That's not a strawman. It's a pretty accurate description of how most casual NotebookLM users operate—myself included, honestly, before I started treating it more deliberately. The output-gets-orphaned problem is real, and it's not unique to NotebookLM. It's endemic to the whole category of AI tools that produce artifacts without a coherent place for those artifacts to live.

Goldie's proposed solution is to wire NotebookLM into what he calls an "Agent OS"—a custom dashboard, built with the help of Claude or another coding agent, that sits NotebookLM alongside three other components: Hermes (an open-source video agent), Hyperframes (an open-source editing layer), and Obsidian (a free local knowledge graph).

What the Five-Layer Stack Actually Does

The framework Goldie calls the "Goldie Knowledge Studio" has five stages, and they're worth unpacking plainly.

Capture: Your NotebookLM notebooks live inside the Agent OS dashboard rather than scattered across Google's interface. You connect them via an open-source MCP (Model Context Protocol), which gives your agents a programmatic way to talk to NotebookLM.

Generate: From within the dashboard, you can produce the full range of NotebookLM outputs—audio overviews, video summaries, infographics, slide decks, mind maps, flashcards, quizzes, data tables, reports—without toggling back and forth between tools.

Sync: Everything generated automatically saves to an Obsidian vault. Obsidian is local-first, free, and crucially, it stores notes as plain markdown files—which means any AI agent can read and write to it. Goldie's setup has agents writing notes into the vault daily, building a persistent knowledge graph that accumulates context over time.

Organize: All assets—notebooks, generated content, agent outputs—live in one dashboard with status tracking. You can see what's in progress, what's done, what's ready to publish.

Publish: Content flows out to wherever it needs to go—social, a CMS, a client folder.

The connective tissue here is Obsidian. "If you use Notebook LM, it doesn't really talk to your other tools," Goldie notes. "But Obsidian links to all of your AI agents." That's the architectural insight driving the whole thing: a shared memory layer that any agent can read from and write to makes the individual tools meaningfully smarter.

Hyperframes: The Part Most People Miss

The piece Goldie flags as the thing "most people completely miss" is Hyperframes—an open-source project that functions as an editing bridge between NotebookLM's generated artifacts and your video agent.

The workflow: NotebookLM generates the raw concept or structure. Hermes takes it and produces a more polished video output. Hyperframes is what lets you edit and iterate on that output without starting over. The optional AI avatar layer—where you inject your own branded persona into the generated video—is Goldie's answer to the "AI content feels generic" objection. Personalization, in this framing, isn't about writing differently; it's about training the system to sound and look like you.

Whether that adequately addresses the genericness problem is a reasonable question to sit with. A system that produces content faster doesn't automatically produce content that's more distinctive. The avatar can look like you. That's not the same as reading like you.

What "Free" Actually Means Here

Goldie is emphatic—repeatedly—that this entire stack costs nothing. NotebookLM is free. Hermes is open-source. Hyperframes is open-source. Obsidian is free. The MCP connector is free. Google's Antigravity CLI (which Goldie notes is replacing the Gemini CLI as of June 18) is free.

This is technically accurate. But "free" in the open-source sense has costs that don't show up in a billing statement: setup time, troubleshooting, the overhead of maintaining a locally hosted dashboard, and the cognitive load of managing a multi-agent system that can break in interesting ways when any single component updates.

"You can ask Hermes to build the agent operating system for free as a dashboard for you," Goldie says. "You just give it the directions. Hey Hermes, create this dashboard for me. Here's how I want it to look."

That's real—coding agents have made self-hosted tooling meaningfully more accessible. But asking an AI to build your infrastructure and asking an AI to maintain your infrastructure are different propositions. For someone comfortable in a terminal who already uses Obsidian, this stack probably does reduce friction. For someone who's never hosted anything locally, the setup cost is non-trivial.

Goldie also offers a paid "AI Profit Boardroom" community where members get the full pre-configured Agent OS, tutorials, a 30-day playbook, weekly coaching calls, and direct access to Goldie himself. The community has 3,100 members, he notes. That's the real product here—and it's worth naming clearly, not as a criticism, but as context. The free workflow exists; the paid option is for people who want the guardrails.

The Self-Improving Claim

One of the stronger architectural arguments Goldie makes is about compounding value. A vanilla NotebookLM setup is static: Google improves the tool, or it doesn't. With the Agent OS, the system theoretically improves as you use it—because your Obsidian vault accumulates more context about who you are, what you've built, and how you work.

"If you go inside Notebook LM, but you don't have a system, and you're just using the defaults, it doesn't get better every time," he argues. "You have to kind of wait for Google to make it better for you. And even then, it doesn't have context on you and who you are and what you do and what you did yesterday."

This is one of the more intellectually honest claims in the video, because it points to a genuine limitation of cloud-hosted AI tools: they're stateless by design. Your NotebookLM doesn't know what you did last Tuesday. Your Obsidian vault, if properly maintained by agents, does.

The question is whether the maintenance burden of that vault stays manageable as it scales, and whether the context it provides is actually meaningful or just noise that grows over time. Goldie's setup depends on agents writing useful notes—which presupposes the agents are writing good notes, which is not always guaranteed.

The Broader Pattern

What Goldie is describing isn't fundamentally new—it's the same problem that every productivity system eventually confronts: how do you get tools that don't talk to each other to share a coherent memory? Zapier built a business on this. Notion tried to be the universal layer. The AI-native version of this problem just has more moving parts and changes faster.

The specific combination of NotebookLM + Hermes + Obsidian + Hyperframes is Goldie's answer to that problem in mid-2025. In six months, some of these tools will have changed, some will have acquired the features that currently require bridges, and the stack will need to be rebuilt.

That's not a reason to avoid building it. It's just the nature of working on the frontier: you're assembling something useful out of pieces that weren't designed to fit together, with the understanding that the puzzle will keep changing shape.

The real question isn't whether this particular configuration is optimal. It's whether the habit of building intentional, interconnected workflows—rather than accumulating isolated AI outputs that go nowhere—is worth developing now.

The orphaned artifact problem isn't going away on its own.

Marcus Chen-Ramirez is a senior technology correspondent at Buzzrag. He covers AI, software development, and the places where technology and power intersect.