Generative AI has already proved useful for accelerating research: summarizing papers, supporting literature review, and producing code snippets. Agentic AI takes this further: agents can access data, write and execute code, inspect results, identify errors, and iteratively refine their approach. For mathematical and statistical research, this means we can explore ideas, run experiments, and test approaches at a pace that wasn't previously possible.

In this talk, we'll explore what this looks like in practice through three examples: experimenting with agentic AI to develop a new approach to estimating π; revisiting an open problem in radial basis function interpolation on spheres; and using agentic AI to help understand, visualise and reproduce results from Elliot Young, Rajen Shah, and Richard Samworth's research on outrigger local polynomial regression.

We'll also discuss what worked, what didn't, and what these experiences suggest about the opportunities and limitations of agentic AI in research.