Summary
In this a16z studio interview, Decagon co-founders walk through the hard-won lessons behind building enterprise-grade AI customer support agents at scale. The conversation opens with a candid account of why Decagon migrated from frontier models like OpenAI and Anthropic to a stack that is now 90% open-source — driven primarily by latency requirements for their voice agent product, and the need to fine-tune smaller models on narrow, well-defined tasks where they can match or outperform much larger general-purpose models.
A centerpiece of the discussion is Decagon’s internal tool called Duet: a second, larger reasoning-focused agent whose job is to automate everything that used to require manual effort — writing agent operating procedures, designing integration tests, running simulations, and monitoring live production conversations for quality regressions. The founders describe Duet as only becoming feasible once reasoning models from Anthropic and OpenAI reached sufficient capability, and frame it as a paradigm shift in how AI products themselves get built and maintained.
The interview also tackles the open-source vs. closed-source debate in enterprise AI, exploring what it means for a startup to ‘own its destiny’ as frontier labs increasingly compete in the same application layer. For engineers and product leaders building on top of AI models, this conversation offers a grounded playbook covering model selection, latency optimization, fine-tuning investment, and the emerging architecture of agents that build and supervise other agents.
📺 Source: a16z · Published July 31, 2026
🏷️ Format: Interview
