What makes an AI product more than a prompt in a polished interface? In this conversation with Greg Isenberg, Gamma co-founder and product leader Jon Noronha answers with the agent harness: the tools and feedback loop that help a model gather context, do the work, inspect it and ask a person when taste matters. Gamma announced that it crossed $100 million in annual recurring revenue in November 2025. The interesting founder lesson here is not that every niche needs a Gamma clone; it is how to find the one job where a specialized loop changes the outcome.
Build a narrow agent that can do a specific job, check its own output and let the buyer steer the subjective decisions. Its advantage must come from a sharper workflow, distinctive taste, permitted data or a real-world service. A list of model calls by itself is not a business.
Watch Greg Isenberg and Jon Noronha
Disclosure and scope: The episode says Anthropic sponsored it and hosted the conversation at Claude Founder House. Jon's account of Gamma's internal model calls, experiments and customer signals comes from this interview; we have not inspected Gamma's systems or reproduced those tests. Gamma's own launch announcement and product guide support the publicly visible workflow.
What the Harness Actually Adds
A model can generate text or an image. A preset workflow can chain calls in a known order. Jon's harness goes further: it gives an agent access to task-specific tools, a place to work, ways to inspect the result and a route back to the user when the brief is incomplete. For coding, that might mean reading files, running tests and changing code. For Gamma, it means understanding the audience, choosing a visual direction, making slides and checking that they are readable.
The useful boundary is not agent versus no agent. It is what decisions must the product make after the first prompt? A ten-slide deck made by a fixed template needs fewer decisions than a presentation assembled from messy notes, a brand guide, connected work material and feedback. A harness earns its complexity when those decisions improve the delivered work.
Inside One Gamma Presentation
Jon walks through a request to turn messy notes into 20 slides. Gamma starts a cloud work environment, reads the files, asks about audience and purpose, can pull in permitted context from tools such as Notion or Slack, drafts an outline, chooses a style and generates text, imagery, charts and layout. A vision-based review then looks for problems such as cut-off faces, bad spacing and weak contrast before the user sees the finished deck. Gamma's public creation guide documents the questions, plan review, generation and targeted editing; its connector guide explains how connected sources can be referenced and what permissions they require.
- Clarify: ask who the deck is for and what story it must tell.
- Plan: outline the narrative and show the style before generating.
- Make: route writing, layout, charts and images to appropriate models and tools.
- Inspect: review the slides visually, fix defects and let the person edit.
By the end of that illustrative run, Jon estimates roughly 20 models and 200 calls. Those are his approximate internal counts, not a published benchmark for every Gamma presentation. The central idea is that the user sees one result while the product handles many small tasks and quality checks behind it.
Gamma describes its October 2026 rebuild as Gamma 5: a more varied visual range, back-and-forth planning, freeform editing and connected context. The short official launch video is useful for seeing the product side of Jon's explanation:
Why Faster Lost the Test
Jon expected that skipping the opening questions and outline would help users reach a finished deck faster. In Gamma's experiment, he says the faster flow lost overall: people who reviewed the story and answered a few questions reported higher satisfaction and upgraded at a higher rate. He also says Gamma retained a faster version for some new mobile users. This is a product-specific result, not proof that every agent should slow people down.
Gamma now documents Plan and Express modes. Plan asks questions and waits for outline approval; Express goes straight to a draft. That makes the experiment legible: good friction is the check that prevents a costly wrong turn. For a high-stakes board deck, audience and story approval may be worth a minute. For a quick mobile sketch, the same questions may be too much.
Taste, Evals and the AI Smell
Jon calls a recognizable, generic AI look or voice AI smell. Gamma's response, according to its launch note, is to broaden styles and let people guide both the plan and the final design. A strict test can check whether a deck follows instructions or uses a brand color. It cannot fully answer whether it feels original, persuasive or like the person who will present it.
Jon describes a mixed scorecard: automated checks for objective requirements; human taste for the subjective result; and product experiments that look at editing, sharing, exporting, paid conversion and later retention. None is a perfect single quality score. Time spent editing, for example, may signal either engagement or repair work. The practical lesson is to pair a measurable defect check with a buyer's judgment of the finished output.
Where a Vertical Agent Can Win
Jon proposes three kinds of edge: a distinct product vision, data others cannot easily access, or a service that combines software with physical work. He also looks for tasks unlike generic coding, where users care about the details and reasonable people disagree on what good looks like. That points to specialized creative workflows rather than another general chatbot.
| Idea from the interview | What a real harness must add | First risk to test |
|---|---|---|
| A Slack collaboration agent | Find actionable threads, learn an approved writing style, draft a reply and leave sending to the person. | Can it respect channel permissions, private context and the difference between a draft and an authorized reply? |
| A couch-shopping assistant | Compare real inventory, render options in a room and check fit against supplied dimensions. | Pixels alone do not establish physical size. Require known measurements or calibrated capture before claiming a couch fits. |
| A professional creative copilot | Learn a practitioner's process, show alternatives, verify objective constraints and preserve manual control. | Does it improve accepted work, or merely generate more attractive first drafts? |
These are ideas discussed in the interview, not Gamma products or validated startup markets. Jon thinks a professional buyer, such as an interior designer or a company team, may pay for a tool that fits their actual work better than a broad consumer novelty. The better starting question is: which task do you know deeply enough to build the missing tools and judge the output?
Models, Prompts and the Cost of Choice
Gamma uses multiple model providers, Jon says, and tests promising releases quickly against small jobs in its harness. He describes the stack as broad but shallow: different models for writing, selection, imagery and review, with much of the advantage coming from context and prompts rather than custom model training. His "98% prompts" figure is a qualitative description of Gamma's approach, not a measured share of value transferable to every startup.
Fine-tuning can come later, he argues, when repeated use reveals a narrow quality problem or a cost-saving opportunity. For a new vertical product, first prove the workflow with available models, track the cost of a finished, accepted result, and test whether routing a cheaper model to a simple subtask changes quality. Switching models is an optimization only when the buyer still gets the outcome.
A First Harness Pilot
My read of the interview is to start with a job that already has a human review point. Interview one buyer, map their current process, then prototype a single loop that produces a deliverable and catches one expensive failure. Compare it with a simple prompt-and-template version. Keep the harness only if it measurably improves reliability, personalization, speed or cost. Jon's closing advice is to choose work you would use yourself; Greg's complementary line is to stay loyal to the niche while changing the product as you learn.
The following is an original editorial brief, not a Gamma prompt or a tested business plan.
One buyer, one job, one feedback loop
Test a narrow workflow before building a platform.
Video Chapters
The harness and Gamma
| Time | Topic |
|---|---|
| 00:00 | Meet Jon Noronha and the agent-harness question |
| 01:23 | What an agent harness is |
| 03:36 | Why the harness is more than a wrapper |
| 04:25 | One prompt through Gamma's workflow |
| 08:23 | Where vertical harnesses fit |
| 11:08 | Three ways to build an edge |
| 11:56 | Physical services and other openings |
| 13:09 | Designing around unique tools |
Quality, models and product judgment
| Time | Topic |
|---|---|
| 13:48 | The problem of AI smell |
| 15:49 | Evals, taste and human judgment |
| 18:40 | Who pays for AI |
| 19:44 | The Gamma for another niche |
| 20:19 | A personal Slack harness |
| 22:01 | Building on other companies' models |
| 23:28 | Closed and open models |
| 24:11 | A broad but shallow model stack |
| 24:57 | Prompts before fine-tuning |
| 25:30 | Testing a new model quickly |
Experiments and marketing
| Time | Topic |
|---|---|
| 26:25 | Gamma's experiment loop |
| 29:10 | The marketing engineer |
| 32:26 | Automations that still need plumbing |
| 33:33 | The top 1% marketing engineer |
| 35:00 | Everyone as a VP of agents |
| 36:19 | Experimentation as a numbers game |
What to build next
| Time | Topic |
|---|---|
| 38:26 | Why image models are underused |
| 40:42 | The couch-shopping harness |
| 45:40 | What Jon would build next |
| 47:26 | Love the niche, revise the product |
Sources and Limits
Checked 29 October 2026. The interview and supplied transcript are the source for Jon's internal pipeline, approximate model-call count, model-selection practice, experiment results and startup ideas. We have not audited Gamma's internal data or built the proposed Slack or furniture products. Gamma's Gamma 5 announcement documents the product direction; its Agent guide explains Plan and Express; the connector guide covers source access; and Gamma's 2025 company announcement is the source for its $100M ARR milestone. Product availability, credits and pricing can change; consult Gamma's current plans before buying.