Alessio Fanelli, founder of Kernel Labs and co-host of the Latent Space podcast, runs two production AI workflows worth studying. The first: OpenAI Symphony wired to Linear, where Linear acts as a state machine and Symphony drives agents through the full dev lifecycle without human babysitting. The second: OpenAI Codex with browser access, autonomously scraping eBay, extracting PSA certificate numbers, and flagging underpriced Pokemon cards in the $10,000 to $20,000 range for his San Carlos retail shop, Merlin Games. These are not demos. They are running in production.
The numbers and architecture decisions are what make this episode worth your time. Fanelli tracks token costs per task and puts a specific figure on the table: 221 million tokens. He explains why local Mac Minis fail at scale and what a cloud VPS unlocks for parallel agent runs. He also makes the case that your CLAUDE.md file probably needs a full purge, not more instructions added to it. The mental model shift from 'agent prompter' to 'agent manager' is introduced at the 2:24 mark and frames everything that follows.
The Pokemon card workflow is the sharpest illustration of a broader argument Fanelli makes near the 24-minute mark: AI is creating a new category of small business that was not viable before. A solo operator can now run autonomous sourcing intelligence on high-value inventory around the clock. The full episode is on YouTube, Spotify, and Apple Podcasts. The Symphony framework is open-source on GitHub at openai/symphony.
[READ ORIGINAL →]