people · May 20, 2026
Princeton SWE-agent Included in A Cross-Framework Behavioral Analysis of Software Engineering Agents
Share the canonical public link.
A new arXiv paper titled A Cross-Framework Behavioral Analysis of Software Engineering Agents appeared 20 hours ago. Researchers analyzed 64,380 SWE-Bench runs across 126 agent configurations from 43 frameworks including SWE-agent and OpenHands. Each configuration paired an LLM such as GPT-4o or Claude 4 Sonnet with a framework supplying tools and workflow. The study separated framework effects from LLM effects by holding layers fixed and measured behavior-outcome correlations.
Below validation threshold — auto-passed without scoring