We cloned our visualization team — They live inside a subagent now
Blog post from Hex
Hex has developed a visualization subagent to enhance its main AI agent's ability to handle visualization tasks independently, improving the quality of visual outputs and allowing the main agent to focus on broader analytical tasks. By offloading visualization-specific tools to this specialized subagent, the main agent is relieved from the complexity and token usage associated with visualization intricacies. The subagent is designed to iteratively refine visualizations by focusing first on foundational aspects such as data structure and axes before moving on to styling. This approach is informed by visualization best practices and the "agent accessibility principle," ensuring the agent perceives and refines its outputs effectively. Evaluation of the subagent's performance is done using a modified version of Microsoft's VisEval benchmark to ensure that it not only executes visualization tasks but also makes qualitative judgments, while bug-specific tests help refine and expand its capabilities. This iterative cycle of development and evaluation aims to continually improve the agent's visualization proficiency, with Hex inviting users and potential employees to engage with and contribute to this evolving technology.