Blog

Notes on how Interaction Bench is built, graded, and what the runs show.

Published on Thursday, October 29 2026

Introducing Interaction Bench: Can an agent fix a bug it has to see first?

Interaction Bench asks coding agents to find and repair usability problems that only show up when a person uses the product. This post covers where it came from, how 44 tasks were built from 4,356 merged pull requests, what the first six agents missed, and how your feedback will shape 1.0 in about a month.