repository-level-tasks

SWE-Together is a multi-turn coding benchmark created from real user-agent interactions, featuring a reactive LLM simulator to evaluate agents based on both final correctness and interaction efficiency.