Coding agent harnesses do not improve task pass rates
Researchers at Shandong University tested Claude Code, OpenCode, and mini-SWE-agent, showing that swapping harnesses yields no better pass rates than a simple rerun.
Resending system prompts with every step triples expenses per task. #coding ⚡ #ai ⚡ #developers
Subscribe
• Sent via TeleFeed
Researchers at Shandong University tested Claude Code, OpenCode, and mini-SWE-agent, showing that swapping harnesses yields no better pass rates than a simple rerun.
Resending system prompts with every step triples expenses per task. #coding ⚡ #ai ⚡ #developers
Subscribe
• Sent via TeleFeed