CharlieKey takeaways from your favorite podcasts
← 29 October 2026The Startup Ideas Podcast

$100M ARR Founder: Build an Agent Harness

John Noronha argues that the agent harness, the tools and loops around a model, is where product value lies. He says to build it where taste and subjectivity matter, test constantly, and stay niche.

Watch on YouTube ↗

Charlie's summary

  1. The agent harness is the real work: 'GPT wrapper' undersells it, since models are similar and the harness makes the product

    John says an agent is an LLM loop that works, checks itself and uses tools, and the harness is the set of tools around it. Claude Code is the best-known example, a harness for coding. Gamma rebuilt itself as a creative harness for visual communication. John objects to the 'GPT wrapper' jab because it implies the LLM does 90% of the work. He thinks the wrapper is the biggest part, and even OpenAI and Anthropic invest heavily in theirs.

    ▶ 1:31
  2. One Gamma request fans out into about 200 model calls across 20 models, ending in a visual quality-control loop

    John walks through a prompt like 'turn this into a 20-slide presentation'. Gamma starts a cloud sandbox, asks clarifying questions, pulls context from Notion, Slack or Linear, and picks a template with vision and embedding models. That is already 30 to 40 calls. Building the deck takes about 100 calls across 10 to 15 models, run in parallel for speed. A visual model then reviews each slide for misaligned text or cropped heads and triggers fixes. The total is roughly 200 calls.

    ▶ 4:31
  3. Build a specialized harness where results are subjective and the user cares about specifics, and ask what your edge is

    John's two signs that a vertical harness is worth building: quality is subjective, so two people may disagree and a human must be involved, and the user cares about the details of the output. He would think about personal games or other creative tools. Your edge can be a distinct view of where the world is going, unique data, or physical-world services that AI can reimagine. Copycats copy where you were three to six months ago, so you must keep differentiating.

    ▶ 8:31
  4. Fight the 'AI smell' by giving people a thousand styles instead of one monolithic look

    John says AI work is increasingly recognizable, in emails and in images, where people can spot a Nano Banana poster from afar. Gamma was a victim of its own success: once everyone used it, everything looked like Gamma. Its relaunch goal was to widen the creative range, so there are many AI smells rather than one. John compares it to fashion, where looks cycle from wide jeans to skinny, so Gamma built an automated system for constantly rotating styles.

    ▶ 13:33

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.