Interactive coding-agent evals measure correction cost

SWE-Together evaluates coding agents in interactive user sessions, using repository-level tasks and measuring both final correctness and how many corrective feedback turns are needed. Why it matters: A coding agent that eventually solves a task may still be expensive if it needs repeated steering. Interactive benchmarks make collaboration overhead visible.

Try this: When comparing coding agents, track not only pass or fail but also correction turns, elapsed time, and whether the same failure pattern repeats across tasks.