Gemini 4 Neon triples SWE-sweep across harnesses
Google's Gemini 4 Neon scored 39.0% on the SWE-sweep benchmark inside the Google Antigravity harness, compared to just 12.8% on a minimal scaffold using identical model weights. The 3x performance gap reveals that context compaction, tool routing, and scaffold design dominate autonomous coding outcomes over raw checkpoint scaling.
The 26-point benchmark swing without altering model weights proves that the primary bottleneck in autonomous software engineering has shifted from foundation model training to agent harness systems architecture.
- –Harness engineering eclipses weight scaling: Scaffolding decisions such as file budgeting, tool execution loops, and error recovery outweigh raw intelligence gains.
- –Benchmark comparability requires rethinking: Without standardized agent harnesses, coding evaluations test scaffolding complexity rather than pure underlying model capability.
- –Systems design guides developer focus: Teams building autonomous coding agents will unlock greater gains optimizing execution loops than waiting for next-generation base models.
- –Antigravity harness demonstrates edge: Google's proprietary agent framework maximizes multi-turn efficiency by combining aggressive context pruning with specialized tool orchestration.
DISCOVERED
1m ago
2026-11-20
PUBLISHED
35m ago
2026-11-20
RELEVANCE
AUTHOR
RahulChaube_