a16z x Decagon: Open-Source Migration and Moat in Enterprise AI Agents

On the a16z podcast, Decagon's co-founders shared engineering and business insights from deploying AI agents in large enterprises like banks and airlines. They revealed an aggressive model migration strategy and argued that engineering and safety capabilities, not models, form the real moat in enterprise AI.

Confirmed

Open-source-first strategy: Decagon runs 90% of inference traffic on open-source models. They recommend starting new use cases with frontier closed models, then migrating to open-source models with fine-tuning once mature.
Small model potential: The co-founders argue that the intelligence-cost tradeoff is a false dichotomy. Even smaller models, after fine-tuning, can outperform frontier models on specific tasks.
Enterprise AI moat: Co-founder Ashwin Sreenivas emphasized that models are never the moat. Even if future base models are perfect, enterprises still need complex safety guardrails, data integration, and engineering to deploy them safely.

Why it matters

As base models rapidly improve and commoditize, business models relying solely on model capabilities become fragile. Decagon's experience shows application-layer companies can build durable moats by embracing open-source and fine-tuning to cut inference costs, while investing in safety guardrails, private data integration, and enterprise-grade engineering.

2026-07-31 ~ 2026-08-01 · 6 related posts

Primary sources

a16z Talks with Decagon: Engineering and Running Enterprise AI Agents at Scale — a16z · 2026-07-31 22:59 (UTC+08:00)
Decagon's Playbook: Why Enterprise AI Agents Are Shifting to 90% Open-Source — a16z · 2026-07-31 22:30 (UTC+08:00)
Decagon Founders: Perfect Models Won't Kill Enterprise AI Moats—Guardrails Will — kimberlywtan · 2026-07-31 23:52 (UTC+08:00)

[source] Decagon's Playbook: Why Enterprise AI Agents Are Shifting to 90% Open-Source — a16z · 2026-07-31
[source] a16z Talks with Decagon: Engineering and Running Enterprise AI Agents at Scale — a16z · 2026-07-31
Decagon Co-founders at a16z: 90% of Enterprise AI Agents Run on Open-Source — kimberlywtan · 2026-07-31
[source] Decagon Founders: Perfect Models Won't Kill Enterprise AI Moats—Guardrails Will — kimberlywtan · 2026-07-31
Decagon CTO: Fine-Tuned Smaller Models Outperform Frontier LLMs on Specific Tasks — kimberlywtan · 2026-08-01

1 near-duplicate retellings: kimberlywtan