TechPrufer
Get started
Hex
Hex logoHex · Feature trail

Visualization Subagent

2 posts · Jun 2024 – Apr 2026

What this trail covers

Hex has evolved its generative data apps by introducing a dedicated visualization subagent. This specialized AI agent handles all visualization-related tasks, taking over from the main Hex agent. The subagent is designed to be iterative, focusing on structure first and then style, and is equipped with enhanced perception capabilities to accurately interpret its own output. It utilizes a structured handoff from the main agent, incorporating principles from visualization design to ensure the 'what' and 'why' are passed, not dictating the 'how'. Evaluation is performed using a rubric modeled after VisEval, with specific tests for bug fixes and proactive expansion of capabilities.

All Hex trails
  1. We cloned our visualization team — They live inside a subagent nowThis post details the creation of a dedicated visualization subagent, separating visualization tools and logic from the main Hex agent. It describes the challenges of generalist agents in handling iterative visualization tasks, the structured handoff mechanism using visualization design principles (what/why vs. how), and the importance of 'agent eyes' (enhanced perception of output). It also outlines the iterative development workflow (structure first, then style) and the evaluation methodology using modified VisEvals for end-to-end and targeted bug fix testing.
  2. Making AI charts go brrrrThis post details the technical efforts to significantly reduce the latency of AI-generated charts in Hex. Key contributions include: 1. Optimizing the Hex chart spec to be more token-efficient (5x fewer lines than Vega-Lite). 2. Creating a 'chart stub' by identifying and using only the 35 most commonly used configuration keys from the chart spec, reducing token count and thus latency. 3. Experimenting with faster models to generate the chart spec. 4. Implementing a system to reliably elicit the chart spec from LLMs, overcoming the challenge that GPT-4 is pretrained on Vega-Lite but not Hex's internal spec. The result is an 8x speedup in AI chart generation, from 10.3s to 1.3s.