Prism Benchmark

An evaluation benchmark for assessing AI coding agents on real-world software engineering tasks. It covers sub-tasks including code generation, debugging, and cross-file modification, using task completion rate and the number of passing test cases as core metrics. The benchmark emphasizes measuring the practical end-to-end usability of agents on authentic engineering workflows.