Gamma CPO says agent harnesses are the new product moat
John Narona of Gamma argues that the tools and context layers surrounding large language models, not the models themselves, create value in AI products. He advises founders to focus on proprietary data and distinct vision rather than training custom models.
John Narona, co-founder and CPO of Gamma, argues that agent harnesses are the primary source of value in AI products, effectively replacing the concept of the 'GPT wrapper.' Gamma, a creative agent for presentations and documents that has reached 100 million in annual recurring revenue, recently relaunched its platform. Narona explains that a single user prompt in Gamma triggers a complex workflow involving a cloud sandbox and 200 distinct LLM calls across 10 to 20 different models. The system also performs automated visual quality assurance to fix layout errors. He asserts that the 'harness' involves curating context, soliciting user preferences, and using vision models to inspect output quality.
Narona advises founders that the competitive moat lies in proprietary data, distinct vision, or non-software integrations, rather than training custom models, which he deems unrealistic for early-stage startups. He notes that Gamma uses 98% prompt engineering, reserving fine-tuning only for cost optimization or specialized tasks where off-the-shelf models fail. He argues that the 'wrapping' or harness is most of the work in building AI products, and that founders should focus on building distinct tools and context layers rather than competing on model training.