Decagon’s Playbook for Building Enterprise AI Applications

Decagon builds enterprise AI by prioritizing open-source models for latency, cost, and control, while utilizing "forward-deployed" engineering to productize workflows rather than relying on permanent manual consulting.

• Model selection — 90% of operational workflows run on open-source models to optimize response times and maintain strict control over model output [00:32].

• Optimization — Engineers tune smaller models for individual tasks, reaching performance levels exceeding general-purpose frontier models while simultaneously reducing operational costs [05:37].

• Architecture — The platform utilizes "glass box" designs where companies retain visibility into agent logic, contrasting with "black box" competitors that mask internal operations [39:51].

• Scaling — An agent called "Duet" automates the maintenance of other agents by writing procedures, running tests, and monitoring performance across millions of interactions [31:46].

• Development model — Forward-deployed engineering serves as a temporary phase to learn operational workflows and build product features, rather than becoming a permanent consulting business [22:51].

• Employment trends — Automation often triggers the Jevons Paradox, where lower costs drive companies to expand the accessibility and volume of support rather than simply cutting headcount [1:17:34].

• How does training smaller models influence operational latency?

• Does implementing autonomous agents typically result in reduced headcount or expanded service offerings?