$100M ARR Founder: Build an Agent Harness

Gamma began in 2020 and took off in 2023, moving from manually chained LLM calls and preset workflows toward an agent. Founder John Narona described an agent harness as the tools and environment that let an agent do work, inspect results, continue across calls, and account for user preferences. He sees the harness as much of the work of making a model into a product, not just a thin wrapper. Gamma’s creative harness must manage storytelling, design, images, layouts, and user taste; a coding harness such as Cloud Code instead gives an agent access to files and tests and rules for planning, testing, and GitHub workflows.

For a request to turn messy notes into a 20-slide presentation, Gamma can create a cloud sandbox containing the prompt, user file, and other context. It may ask clarifying questions, retrieve information from Notion, Slack, or Linear, draft an outline, and choose and customize a template before generating slides. Narona said those early stages can require 30–40 LLM calls across four or five models. Text, layouts, images, diagrams, infographics, and charts add further calls, often run in parallel to reduce latency. A deck may start streaming after about a minute; one estimate put the full process near 100 calls and 10–15 models, while a separate estimate described roughly 200 calls and 20 models for a finished creative result.

Gamma treats the first draft as unfinished: vision models check for issues including misaligned text, cropped heads, contrast, and padding, and the system can run several rounds of checks and fixes. The host said he thought Gamma had reached $100 million ARR and had recently relaunched its platform to offer a wider range of generated styles instead of one recognizable Gamma look. Narona compared changing styles to fashion trends and said the goal includes reducing the “AI smell” that makes generated writing or imagery feel recognizably machine-made. Objective evals can test instruction-following or adherence to brand guidelines, but are less able to judge whether a design is cool or sounds like a particular user.

Because creative quality is subjective and users differ, Gamma also uses large-scale A/B tests of prompts and models. It tracks editing time, sharing or exporting, continued use, and paid conversion, which Narona described as a favored North Star because it is hard to fake. Gamma generally tests promising models within an hour, using evals and samples of live users, and often runs 10–20 experiments at once, many concluding in under a day. After Claude Haiku 5.5 was released, it started eight tests, including slide edits and outline revisions. Some tests show a model costing half as much performs equally well, while other tasks still need a larger model. Skipping outline review and upfront questions hurt satisfaction and paid conversion overall, though Gamma shipped a fast-mode version for new mobile users, who may find typing answers inconvenient.

Narona advised early-stage companies to use existing models rather than assume that training or fine-tuning is necessary for a moat. After real usage develops, teams can fine-tune to lower costs or improve personalization, or build specialized models where frontier models perform poorly. He called closed-source cloud models underrated and estimated that prompts account for 98% of his team’s AI value, with fine-tuning mainly useful for cost optimization or the final 2–3% of quality. A defensible business may instead come from distinct vision and fast execution, proprietary data, domain-specific taste, or combining AI with hardware, physical services, or other real-world operations. He saw legal-tech AI as crowded, while medical technology may benefit from data lock-in and regulatory complexity.

The discussion also pointed to opportunities in work-focused agents and specialist workflows. Narona sees professional AI spending rising quickly and work subscriptions as more attractive than consumer subscriptions. A Slack harness that sorts messages, drafts replies, and learns a person’s voice from observed actions could help, but authentic representation in professional settings remains difficult. Gamma users can turn Salesforce opportunities and call notes into pitch decks, though connecting the workflow may still require tools such as N8N or Zappi; the broader opportunity is making specialist-built workflows usable by any business. The speakers also described marketing engineers using agents to run many SEO, ad, social, and brand experiments, while anticipating that investment in brand identity may counterbalance performance marketing at scale.

Narona sees image models as useful beyond familiar image generation: Gamma uses about 10 for tasks such as diagrams, infographics, textures, and custom icons. A harness could measure whether a couch rendered in a room matches exact dimensions and prompt corrections, though he doubts model labs will build that workflow soon. Furniture visualization and interior-design products could serve consumers and professionals, with professional users and larger companies looking like stronger prospects for durable businesses. More broadly, he sees vertical AI and SaaS opportunities that can start in one niche and expand; Gamma earns less than a quarter of its revenue in the United States and serves consultants, teachers, marketers, salespeople, and lawyers. His startup rule is to “marry the niche, date the product”: stay committed to the people served while testing and iterating on ways to serve them.