A multi-agent system built solo on the Claude Agent SDK. It scouts the day's AI news, fact-checks itself, narrates an episode and emails it. It runs free, on a schedule, with no server.
AI moves faster than anyone can keep up with, and the people who most need to stay current have the least time to read. The gap isn't access to news. It's time. So I designed around the one slot left in the day: the commute. Eyes busy, hands busy, twenty-odd minutes, already listening to something. That single constraint set the format (audio), the length (short) and the tone (a podcast, not a wall of bullets).
A team of agents modeled on a real newsroom, built on the Claude Agent SDK with Haiku as the reasoning model to keep daily cost near zero.
What makes this a real agentic AI system and not a pipeline wearing the label: the agents make decisions, they call tools, and there's a critique-and-revise loop. The Fact-checker can reject the Writer's work and demand a second pass. That loop is what makes the output trustworthy enough to actually send.
It produces a real artifact every day. A formatted brief plus a narrated audio episode, delivered by email and scheduled free on GitHub Actions.
The AI was the easy part. Scout, Editor, Writer and Fact-checker worked early and worked well. Every real fight was in the glue around the model.
The pattern is the lesson. Almost none of these were the agent logic. The model was the easy twenty percent. The plumbing was the other eighty: cloud quirks, silent failures, file formats. That's the real story of shipping an AI product, and it's the part tutorials never show.
A daily AI brief is a crowded category, and being honest about that is part of the work. In May 2026 Spotify shipped the generic version: personalized daily audio briefings, with distribution no solo project can match. That could have killed the idea. Instead it sharpened it.
Once a platform commoditizes the generic version of a feature, you don't win on the generic version. You win by going narrower and deeper than a horizontal tool ever will. I also named the assumption most likely to kill the product, that people will give up their commute music for a work brief, and designed a small two-week pilot with real users to check it, with the success and failure criteria set in advance so I couldn't move the goalposts.
The sharpest lesson was about where to be brave. In a crowded market the brave move isn't a flashier feature. It's narrowing hard enough that a horizontal tool can't follow, and being willing to let a small pilot kill the idea before you over-invest.
What's next: run that pilot for real. Pick one profession and go deep. Publish the podcast where people already listen: Spotify, Apple Music, YouTube. And A/B test what actually drives listening: length, voice, send time.