I analyzed 246 repos and 57 papers on agent harnesses. Here's what actually works

An analysis of 246 repositories and 57 papers on agent harnesses reveals that many developers are building these systems without measuring their effectiveness. A key finding suggests that limiting the tools and skills exposed to an agent significantly improves performance. For instance, Vercel saw success rates jump from 80% to 100% and token usage halved after removing 80% of their tools, while Microsoft reduced 100 tools to just 2, demonstrating that fewer tools can lead to better results.

This report uniquely synthesizes findings from 246 repositories and 57 papers, offering a rare, data-backed perspective on agent harness effectiveness, unlike most anecdotal accounts.