Using AI Agents to Create Online Dashboards
For hands-on training in designing structured AI agent workflows with built-in checks, human oversight, and clear documentation, join Professor Crabtree for AI Agent-Driven Research Workflows.
The introduction of AI agents has brought major changes to higher education and promises many more to come. Some have argued that the days are numbered for old ways of producing research, and the same could be true for old ways of disseminating research. In this post, I’ll highlight one possible option for academics who want to move past the paper manuscript as the main (or only) unit of research output. I’ll talk about online dashboards, how they might help researchers, and how they can use AI agents to build them.
To ground ideas, I define a “dashboard” as an interactive page that lets a user examine, filter, sort, describe, and analyze existing data or results from a project. A dashboard doesn’t need to be complex. It could be as simple as a self-contained HTML file, a Shiny app, or even a Quarto document.
The idea of using dashboards to highlight research findings is not new. For instance, Pablo Bernabue suggested in 2017 that “alongside files and code shared in repositories, …[researchers could] present the data in a website, in the form of plots or tables”. Even earlier in 2011, Bret Victor advocated creating “reactive documents” and “explorable explanations”. These and other authors have argued that dashboards can help broaden research impact by improving understanding, inviting user discovery, and enabling others to extend research.
While the benefits seem clear, until recently the practical costs of creating interactive dashboards have prevented their widespread adoption. Before AI agents, a veteran programmer might have taken a week or more to build a suitable dashboard using Shiny or D3. While I have long believed in the value of interactive dashboards and am an experienced R programmer, I rarely had time to build dashboards to accompany my own research. Thankfully, the same task now takes a couple of hours at best and an afternoon at worst.
For example, I recently built a dashboard to highlight Ron Francisco’s extensive protest data. Francisco spent years creating daily records of protest and state coercion, compiling tens of thousands of events between 1980 and 1995 across 28 European countries. While he made these data publicly available on his faculty site, their format made them difficult to use. To highlight his major data contribution, I built a dashboard that lets users interact with the cleaned, compiled data: they can now view summary statistics, explore annual trends, search all events, and run their own regression models. In all, what may have taken hours of work to do manually took less than an hour using AI.
It’s easy to have an AI agent build something like this. The figure below outlines the steps and the animated video walks you through them. First, you point the agent at your data, tell it that you want to create a dashboard, and define its functionality and audience. Second, you specify the output format. There are lots of options, differing in both how complicated they are to implement and maintain, and what kind of interactivity they provide. I prefer to specify a self-contained HTML-based dashboard. Compared with other options, I think that it’s easier to develop, tweak (with or without an agent), share, and post online. Third, you’ll want to test the dashboard repeatedly, checking both functionality and outputs. Fourth, you’ll probably want to iterate through several user interface designs until you’re happy with the final product. As an optional step, you might ask the agent to create a reusable skill that encodes your stylistic preferences once you’ve finished creating the dashboard. This way you can invoke the skill for new dashboards and simplify the design process. You can find all of these steps in a handy diagram here:
Once the dashboard has been created, you can share it on your own website or on a hosting service like GitHub Pages or Vercel. You could also register the dashboard with services like Figshare and Zenodo, which would provide your dashboard with a DOI.
In my seminar AI Agent-Driven Research Workflows, I’ll show you how to build dashboards both to broaden research impact and to validate AI agent outputs. I’ll also teach you how to develop other agentic skills that will increase your efficiency and the quality of your work. No prior experience with AI agents is assumed.
Hope to see you there! And good luck creating dashboards for your research!
AI disclosure statement: Claude Code was used to brainstorm, identify accessible internet resources on this topic, develop an initial outline, and polish the writing. Claude Design created the animation and figure. Grammarly was used for spelling and grammar assistance.