Terence Tao Builds Apps With Coding Agents

This research library uses AI-assisted source research and drafting. Linked sources support product claims; analysis and proposed exercises are our interpretation. Unless an article documents a test and its results, do not read it as a hands-on review or an independently verified benchmark.

In Old and new apps, via modern coding agents, Terence Tao describes restoring old Java applets as JavaScript applications and creating new mathematical visualizations. He reports a small dragging bug in the migrated collection and explains that these visual aids are secondary to the mathematical argument. Those are his observations about these projects, not a benchmark of coding agents generally.

Updated 21 September 2026: combined two accounts of the same post, removed ranking language and unverified forum reactions, and separated our proposed evaluation from Tao's reported experience.

Write the mathematical contract first

The following is our suggested exercise for a small mathematical app. Choose a visualization whose behavior you can judge independently: a recurrence, a transformation of a shape, or a plot with a few controls. State the formula, parameter domain and known examples before asking an agent to build it.

Do not use the generated implementation as the only source of expected values. Work out a few cases separately and include an edge case, such as a zero input or a boundary parameter. If the model generates both the computation and its expected output from the same mistaken assumption, the test can pass without establishing the intended behavior.

Keep the first artifact in one experiment folder. Name the permitted files and existing dependencies. These are task instructions; enforce any necessary access restrictions in the environment rather than assuming a folder named in a prompt is a sandbox.

Inspect the interaction as well as the formula

A correct expression can still be presented incorrectly. Test labels, units, scales, rounding and the relationship between a control and the displayed value. Drag outside a drawing area, resize the window and reset the inputs. Inspect both the picture and the numbers that produced it.

Ask for a screenshot or rendered check of the actual app. A screenshot establishes a particular visible state, not every possible interaction. Keep a small set of repeatable checks for the important controls, and record any behavior you only inspected manually.

The Tao post makes the distinction between a useful visual aid and the core argument explicit. Apply that distinction in the app's own explanation: a convincing animation is not a proof of the claim it illustrates.

Decide whether to retain the experiment

Use a short review record rather than a ranking of agent products:

Run the documented startup command from a clean state before handing the app to another person. Check that it does not depend on an unrecorded local file or a development server left running from the agent session.

Keep a successful artifact as a teaching aid or disposable experiment when that meets the need. Moving it into a maintained application creates additional requirements around accessibility, compatibility and ongoing correctness. That should be a separate decision supported by the intended users and failure consequences, not an automatic reward for an impressive demo.