#012: My First AI Agent is Alive—Meet Brokaw
I finally stitched together my first AI agent. I named it Brokaw, after the no-nonsense news reporter Tom Brokaw. It automatically feeds me a daily intel brief of everything I'm curious about.
I’m sick of hearing everyone talk about AI agents because I’ve never used one or seen one used for anything practical. Mostly, I’ve just been chatting with AI assistants, not ‘agents’. So the past few weeks I’ve been on a mission to figure this agentic AI thing out.
My essay last week was all about configuring NemoClaw on my DGX Spark. This is really just a combination of Nvidia’s LLM called Nemotron and OpenClaw as the agentic AI system that connects tools and actions. I thought it would be easy to set up, I was wrong. It’s been pretty painful but I made it through.
As I write this, my first AI agent is running (sort of) in the background. It’s called Brokaw. I named it after the no-nonsense news reporter Tom Brokaw, because that’s all I want this agent to do—ship me a daily report of events around the world on topics I care about. It’s like a personalized President’s Daily Brief. I will continue to tailor it over time with all my favorite podcasts and sources. For now, it’s basic.
But the cool thing is that I get to customize the UI and it’s all done locally, except for the part where it scans news sources for the latest updates. And as a next step I will create a Slack integration so that I can interact with items of interest and have it appear on my phone.
I know that this agent isn’t groundbreaking, but I needed prove how this works before I expand into my ultimate vision of creating a workforce of digital employees.
Here’s how it works:
Inference: This is all done locally on my DGX Spark which has a Blackwell chip on it and can provide the GPUs necessary to power an LLM provided by Nvidia called Nemotron. It’s a 49B parameter Llama-based model customized by Nvidia, that isn’t as good as OpenAI or Claude, but it gets the job done locally, so I like it. The less data I feed to OpenAI or Anthropic, the better.
Agentic AI OS: I use OpenClaw on my DGX which like an agentic AI operating system that has a set of tools baked in to allow for agentic actions. When combining OpenClaw with Nemotron, this is called NemoClaw. It was pushed heavily at the Nvidia GTC Conference this week for enterprises as the “Agentic OS”. I am using it to build my personal agentic OS on my DGX.
User Interface: I used Cursor on my Mac Studio to code a personal assistant app accessed through a browser, which will contain all of my future agents. Brokaw is basically a module in my personal assistant app that I can design and configure through Cursor. I simply don’t have the time to use the DGX Terminals to manually edit .js files. Too painful. So Cursor is a natural language way to build at rapid speed.
Prompts: I mainly used Claude Opus 4.6 to create my Cursor prompts. Then I copy errors I get in Cursor or in my UI, tell Claude, then Claude helps me fix them. This is how I keep Cursor honest. I then paste Claude’s guidance back into Cursor, which is a powerful combination. You might be wondering why I don’t just use Claude Code directly, and mainly it’s because of their usage limits and I don’t want to pay $200 p/m! Cursor has handled everything I can throw at it over the past year, so if it’s not broke, I’m not fixing it.
After I generated a briefing successfully, I began celebrating, thinking I had an AI agent configured. But then I asked Claude, “Is this really an AI agent?” Claude had the audacity to tell me it wasn’t—it was just a ‘pipeline’. So I continued to prompt and build, and then I added agentic functionality to Brokaw.
Here’s the “agentic” process:
[Brokaw] Pass 1 — draft complete
[Brokaw] Pass 2 — evaluate & improve: calling LLM…
[Brokaw] Pass 2 — done. Parsed quality: N/5
[Brokaw] Run: saving final briefing (quality N/5)
I asked Claude why Brokaw is now considered an agent, and to explain it to me like I was dumb, and here’s the response:
Brokaw is an agent because it does the whole job on its own. It wakes up at 7am, goes and grabs the news, decides what’s stale and throws it out, writes a briefing, reviews its own work, fixes mistakes, and saves the final version — all while you’re still asleep. Nobody tells it what to do step by step. It figures it out, handles problems when things break, and delivers a finished product. That’s what makes it an agent instead of just a chatbot waiting for you to type something.
I do think this will be worth it. But given how complex it’s proven to be, there’s definitely a lot of hype around “agentic AI” these days. It certainly isn’t something you can just pick up and run with. Brokaw broke so many times this past week, that I was thinking of calling it BrokeClaw.
Anyway, more to come as I think through and build my next agent on this quest to build an agentic AI workforce that works for me in the background. It’s still to be determined if this is real or hype, we shall see. If nothing else, it allowed me to create this fancy video with Grok Imagine, enjoy: