Report
Google paper: Gated agent-harness edits nearly match full rewrites with roughly 90% fewer tokens
DAIR.AI reports gated module edits achieved 40% accuracy versus full rewrites’ 42.5% in a 40-task test.
TLDR
The token bill is the twist: DAIR.AI reports that gated edits to an agent’s harness—the code managing context, tools and checks—solved 40% of 40 held-out SWE-bench Lite tasks at 46.7k tokens per task. Full rewrites reached 42.5% at 451.6k tokens. Gemini 3.8 Flash stayed frozen; modular edits required validation under a token cap. Promising efficiency, but 40 tasks across two repositories is a small test, not a sweeping verdict.
Combined views
13.6K
4 Sources, first seen ago

Comments