Coding Agent Performance varies by Context Management Harness

Changing how a coding agent manages context can significantly impact performance. Under a 20,480-token window, a treatment that shortens older tool results increased complete solutions on SWE-bench Verified from 43% to 72%, despite increasing the fail-to-pass fraction.