SWE-bench Insights: Evaluating Coding Agents on Real GitHub Issues

How SWE-bench changes expectations for coding assistants by grounding evaluation in real repositories and fixes.