How Decagon Runs 90% of Its Agents on Open-Source Models

Summary

Sarah Wang and Kimberly Tan interview Decagon co-founders Jesse Zhang and Ashwin Sreenivas about building enterprise AI agents. The conversation covers Decagon's shift to open-source models, fine-tuning and evaluation, enterprise deployment, the application-layer versus frontier-lab debate, CRM's role as a system of record, and AI's impact on customer support jobs. Key market implications include durability for application-layer software, continued value in CRM systems, and a unique forward-deployed model at Palantir.

• Decagon runs 90% of workflows on open-source models, using fine-tuned small models for production tasks.
• Fine-tuned small models can outperform frontier models on specific tasks while being faster and cheaper.
• The application layer is expected to remain durable even as frontier labs add capabilities.
• Decagon's product vision expands from customer support to a proactive AI concierge.
• CRM systems are seen as valuable systems of record that AI agents will use more.
• Palantir's forward-deployed model is described as unique and able to close massive enterprise deals.
• AI may automate mundane support work but not eliminate careers, with potential Jevons paradox effects.