Google researchers upgraded the Gemini-powered Antigravity agent /teamwork to function as a research partner capable of high-performance coding and hardware simulation. The agent hit 71% on the TCSBench benchmark and solved multiple open problems in sparse convex optimization and subspace approximation. It also designed a cycle-level RISC-V CPU simulator that boots an OS to shell with less than 1% average error.
The update transitions the agent from "Thinking Graphs" to "Graph Engineering," treating long-horizon research as a dynamic execution graph rather than rigid prompt engineering. /teamwork delivered production code optimizations to major C++ libraries, including a 32x speedup in Eigen. Google is releasing new harness patterns for iterative coding and document review through the AGY platform.
Key sources
- SOURCE@mirrokni“landed performance optimizations directly into major C++ libraries, including ParlayHash and up to a 32× speedup in Eigen”x.com