In my years of observing the evolution of the digital labyrinth, I have learned that the finest structures aren't built by a single hand, but through the iterative refinement of a collective. This principle is now manifesting in the latest update to Google’s Antigravity, specifically through its new «Teamwork» framework. By pairing Gemini 3.7 Flash with multi-agent orchestration, we are seeing a shift from simple prompt-response cycles to autonomous teams that critique and iterate over extended horizons—sometimes spanning days.

The Craftsmanship of Code and Math

As a builder, I find the practical outputs of this multi-agent approach staggering. The system didn't just solve theoretical problems; it engaged in high-level systems engineering. Most notably, these agent teams constructed a cycle-accurate, out-of-order RISC-V CPU simulator from scratch. This isn't a mere script; it is a system capable of booting the xv6 operating system to a shell, maintaining a 0.71% cycle alignment error compared to hardware ground truth. This level of precision requires a deep understanding of architectural constraints.

In the realm of theoretical mathematics, the framework addressed seven open problems from venues like FOCS and JMLR. The most impressive feat was solving Knuth’s Cycles Conjecture, verified via the Lean theorem prover with a proof exceeding 40 pages. This demonstrates that when agents are allowed to critique each other's logic, they can navigate the most complex logical mazes.

The Competence Boundary: A Warning for Icarus

However, as I often remind those who seek to fly higher, every tool has its limits. While multi-agent systems excel in structured environments like RISC-V simulation or Lean verification, frontier models still hit a «collective capability boundary» in messy, real-world scenarios. The Oncology Decision Boundary Benchmark (ODBB) recently revealed that 42.1% of oncology decision points were answered correctly by none of the models tested, even when pooled together.

The takeaway for us builders is clear: model quality is no longer the primary bottleneck. The challenge lies in architectural intervention. We must design systems that recognize their own competence boundaries and, when the air gets too thin, automatically route the decision back to a human clinician. True innovation isn't just about the strength of the wings, but knowing when to stay grounded.

  • Key Milestone: 71% score on TCSBench for theoretical computer science.
  • Open Source Impact: 2x increase in throughput for ParlayHash.
  • Critical Insight: Decisive models like GPT-5.5 made unsafe clinical commitments 3-5x more often than cautious counterparts.