Four founders building AI-powered products share lessons from shipping agentic systems at scale. Key themes include: the model itself is no longer the differentiator — reliability is; human oversight remains essential even as agents improve; model selection involves trading off cost, latency, intelligence, and capacity; true 'agency' in LLMs is a misnomer since they require constant prompting; and the real moat comes from solving genuine problems, accumulating proprietary data, and grinding on reliability over time. Founders from Andi AI, Higgsfield AI, LawVo, and Probably discuss what they underestimated in production, how they choose models, and where competitive advantage actually lives when everyone accesses the same frontier models.
Table of contents
Making agents work in productionModel selection is a four-variable problemThe moat is in the execution132 Impressions