The LLMs Are About the Same—What’s Different Is the Layer of Tools Around Them
The capability gap between LLMs is narrowing, while the product gap is widening. Jon Noronha, co-founder of Gamma, says: packaging is most of the work. This piece breaks down what an agent harness is, how to choose a niche, and who is willing to pay for it.
Greg, the host of The Startup Ideas Podcast, offered a judgment: the agent harness is the new AI wrapper. Guest Jon Noronha, co-founder of Gamma, agreed and took it a step further: “Packaging is most of the work.”
The capability gap between LLMs is narrowing. The product gap is widening.
An agent harness is the set of tools an agent uses to complete a task. The agent is the high-level LLM loop that does the work, checks it, calls tools, and turns the result into a finished product. The harness is all the tools this loop uses to act in the environment, check the work, and ask the user what they want.
Claude Code is the harness for programming. Gamma is the harness for visual communication: it learns your artistic style, helps you structure your story, and checks whether each slide is appropriate.
Gamma’s process for running a 20-page deck is: Prompt → cloud sandbox → questions → connectors → outline → template → generation → visual QA → fixes → final product. It involves about 20 models and 200 calls per run; the user sees the final output.
There is a popular saying: there isn’t much room left for vertical AI products. Greg and Jon see the opposite. The opportunity comes from the needs of people doing specific work. What kind of work is suited to a vertical harness? Jon looks for two traits: the work involves judgment—two people may disagree on what 'good' means, so human involvement is needed; and the users care about the specific details of the output. Taste is the edge. In fashion design, he wants a hundred different harnesses—one for Nike, another for Adidas.
Start with a domain you understand. Asked which “Gamma for X” he would build, Jon pointed to founders’ own circles: Pokémon cards, fashion, sneakers.
Before you start, answer: What is my edge? Jon gives three directions:
- You have a unique vision and can move toward a direction faster than others.
- You have proprietary data and information others don’t have.
- You do things beyond software, redefining traditional industries with hardware or physical services.
Speed protects the first kind of edge. Greg says copycats only copy the state you were in three to six months ago. His countermeasure is a product-studio mindset: ship a new zero-to-one feature every two weeks.
Crowdedness also matters. Legal tech already has many harnesses; Greg wouldn’t touch it today. Jon suspects healthcare is the next big opportunity, with both data lock-in and regulatory barriers in place.
Tool design starts with one question: What unique tool do you already have, or what tool could you build that the market doesn’t have? Context engineering is “pretty much the whole game,” broken into three things: bring the right context in from the user; present AI output to the user in the right way; and break down the work so that every LLM task becomes easier.
“When you make the LLM’s work very easy, any off-the-shelf model can give you good results.”
98% of Gamma’s value comes from the prompt. The real battlefield is actual usage. Once you have usage, then fine-tune to lower costs and personalize. Only specific tasks that big models do poorly are worth training your own model for—like Gamma’s slide QA.
Consumers hate subscriptions. Jon admits he’s still mooching off someone else’s Netflix. Work is different; the boom in subscription SaaS has already proven this path can last ten years. Jon uses sofas to segment buyers:
- Consumers: want someone to help them buy a sofa.
- Prosumers: part-time interior designers who charge by the hour and are willing to pay a subscription for better tools.
- True professionals: employees at Restoration Hardware.
One niche can grow into many. Gamma’s users range from consultants and teachers to marketers, salespeople, and lawyers; less than a quarter of revenue comes from outside the U.S. Start with one small niche, then look for adjacent areas.
“A big company is a thousand small niches stacked together.”
On the show, Jon and Greg offered several open harness ideas:
- Slack: observe how you reply, then draft the first reply in any thread that mentions you. Jon says he has sent “about a billion Slack messages.”
- Furniture: render a sofa into a buyer’s room, measure it pixel by pixel with ordinary code, then prompt a model to adjust the dimensions. A 68-inch sofa and a 58-inch sofa don’t look the same against a wall. Jon doubts big model labs will build this anytime soon. Greg adds that sofas have high ticket prices and high margins.
- Personal video games: coding agents can already make games. The opportunity is a harness that asks players where they want the game to go, built for non-programmers, possibly running in a 3D world.
- Marketing automation: the top 2% of companies have used n8n or Zapier to build workflows like “new Salesforce deal → generate a pitch deck from call notes.” Turn that into something every company can configure.
- Service businesses: “boring businesses” like HVAC companies.
- Personal agents: tools that help ordinary people act as the VP of their own lives; Jon thinks this is still underexplored.
- AI video: when everyone can make video, the startup opportunity above video comes back to the harness.
Action checklist:
- Write your harness in one sentence: the tools an agent needs to complete a task.
- Choose a niche that is full of judgment and where users care about the output.
- Name your edge: vision, data, or something beyond software.
- List the unique tools you already have or could build.
- Bring in the right context and make every LLM task simpler.
- Sell to prosumers and professionals, then expand into adjacent niches.
- Put up a landing page and test the niche.
Greg closed with a question: In a world where anyone can make video, what startup would you create? The answer still comes back to the harness.
“It all comes back to the harness, baby.”
Jon’s startup principle is: build a harness you yourself would use—one you can use every day and genuinely believe in. Greg puts it more bluntly: marry the niche, date the product.
Jon Noronha said at the end of the show: “It’s a long and sometimes lonely journey. That’s exactly why you have to truly love what you do.”
发布时间: 2026-11-22 17:45