Speak at The Fifth Elephant 2026 Annual Conference

Describe your session in 2 paragraphs

Generative UI changes what an agent can return. Instead of answering with a wall of text, the agent can render charts, forms, workflows, tables, dashboards, and other interactive interfaces. That looks great in demos, but production exposed a harder question for us: what should the model actually emit to make this reliable, fast, and streamable?

At Thesys, our first version used JSON. It was the obvious choice, and it worked well enough early on. But once we had 10,000+ developers building with it and real traffic going through the system, JSON started showing its limits. Outputs were token-heavy, latency grew with response size, and partial or malformed JSON during streaming often meant retries or broken renders. The core issue was simple: JSON needs to be structurally complete before it can be parsed, while streaming UI needs to become useful before the full response is done.

This talk is about the engineering path we took after that realization. We replaced JSON with a compact, line-oriented intermediate language designed specifically for streaming UI. I’ll walk through the production constraints that shaped the system: a swappable design system, a swappable model layer, a swappable renderer, and a neutral format sitting between them. We’ll cover the streaming parser, component registration, constrained generation, failure modes we saw in production, and benchmarks showing 50 to 67% fewer tokens and 2 to 3x faster render latency. The session ends with a live side-by-side demo comparing the old and new approach.

Mention 1 to 2 takeaways from your session

1. The format an agent emits is not an implementation detail. It directly affects cost, latency, reliability, and how well the product can stream.
2. A neutral intermediate language between the model, design system, and renderer gives teams a cleaner abstraction and agnostic architecture.

Which audiences is your session going to be beneficial for?

This session is for engineers building AI agents, LLM-powered products.

Bio

Zahle is part of the founding engineering team at Thesys. Previously he was part of platform team at Razorpay

Link to draft slides

Slides