Stop building AI wrappers. Start building agent harnesses.

Gamma’s central lesson is that useful AI products require an agent harness, not just a wrapper around a model. John Narona said Gamma began in 2020, took off in 2023, and rebuilt its product this year from one-off LLM calls and preset workflows into an agent. A harness gives an agent tools and context to work in an environment, inspect its output, learn user preferences, and deliver a finished result. Coding harnesses help agents inspect files, run tests, and follow rules; Gamma’s creative harness must also support style, storytelling, image generation, and visual layout.

A presentation request can involve clarifying questions, context from Notion, Slack, or Linear, outline creation, and template selection using vision models and search embeddings. Gamma may make 30 or 40 LLM calls across four or five models before generation begins, and nearly 100 calls across 10 to 15 models over the full process. Parallel generation lets a presentation stream in about a minute. Afterward, a vision model checks each slide for issues such as misaligned text, cropped heads, contrast, and padding. Narona said available models have performed poorly at slide QA, leading Gamma to research models for that specialized task.

Gamma’s AI stack is broad but shallow: it uses many models, often with straightforward prompting and context engineering. Narona estimated prompting delivers about 98% of the value for Gamma; fine-tuning is mainly useful for lowering costs or making smaller quality gains. The team tests new models quickly and assigns different sizes to different tasks, often running 10 to 20 experiments at once, many lasting less than a day. After Claude Haiku 5.5 came out, Narona started eight tests, including slide micro-edits and outline revisions. Tests can show that a model costing half as much performs just as well, or that a more expensive model is needed.

Gamma combines objective evaluations with testing on user samples for subjective work. Evals can measure whether an agent asks clarifying questions or follows brand guidelines, but there is no single objective test for whether creative work feels generic or captures a personal voice. Large-scale A/B tests compare models and prompts across Gamma’s diverse user base, while the company tracks editing time, sharing or exporting, paid upgrades, and retention. Narona called paid upgrades a favorite North Star signal because they are hard to fake. A fast-mode test also showed that fewer steps do not always improve the experience: users who reviewed outlines and answered questions reported greater satisfaction and converted to paid plans at a higher rate. Gamma still shipped fast mode for new mobile users, for whom typing answers may be undesirable.

Gamma broadened its creative range after users began recognizing its early output as distinctly Gamma and did not want their work identified as AI-generated. It built an automated system to cycle through different creative looks, a process Narona compared to changing fashion trends. The speakers advised founders to build harnesses where tasks involve subjective judgment, human input, or close control over results, and to seek an edge in product vision, proprietary data, or combining AI with hardware or physical services. Startups can generally begin with existing models rather than pay to train their own, then focus on real usage and consider fine-tuning to cut costs or personalize results. Specialized models make sense when existing ones perform poorly.

The speakers urged founders to work in domains they understand from experience and to “marry the niche, date the product”: remain committed to the people being served while changing the product as needed. They described a possible Slack harness that learns how a user communicates from their actions and messages, rather than from a quiz; one speaker cited roughly one billion Slack messages as potential examples, while acknowledging that authentic personalization remains difficult. A couch-shopping harness could render furniture in a customer’s room, measure its apparent size programmatically, and correct inaccurate images; dimensions matter, and image models do not guarantee exact scale. Furniture’s high prices and margins could support retailer commissions or paid tools for interior designers. The speakers saw professional-user products as potentially more durable than broad consumer tools.

The conversation also anticipated a “marketing engineer” era in which agents help marketers run many daily experiments across SEO, GEO, social content, paid ads, and brand creative. More creation by agents could shift people toward supervisory roles, though overseeing layers of agents may create information overload. Narona cautioned that personalized websites and apps have not broadly become the norm, and resistance to personalized or surge pricing, along with investment in brands, may limit highly optimized marketing. More broadly, startup building depends on repeated experiments, not one decisive test: early attempts often fail, so founders need to keep going and can begin with something as simple as a landing page or a testable product.