Some questions are answered by a shape: where the time went, which week things turned, how two numbers move together. An agent can draw a chart under its reply, with whatever data it gathered: bars, lines, a scatter you can zoom into, a heatmap, or something no template has, since it writes the drawing itself. Hover, zoom and tooltips all work inside the chart.
Charts take on the app's light or dark look, fit whatever width they're shown at, phone included, and redraw when that changes.
In the desktop app, the agent sees its chart before you do. Frontier draws each one out of sight first, and a chart that breaks, comes out empty or never finishes goes straight back to the agent to fix in the same run. It can also ask for a small picture of the chart, to check for overlapping labels or a clipped axis. If one still fails in front of you, the error shows in its place with "Ask the agent to fix", which sends it back into the thread.
A chart can run things too. Click a bar and it can open that item or run a capability for it, but only the ones the agent named when it drew the chart, and only after you allow each one in a box that shows exactly what will run. Like any reply you can use, a chart can follow a saved value and redraw as it changes.
It's safe for the agent to write that code because the chart runs sealed off: no internet, no access to anything else in Frontier, nothing but the data it was given. Turn your data into a chart walks through asking for one.