Back to feed

Advancing science & engineering with Antigravity’s /teamwork

The latest research from Google

Nov 13, 2026

11/13/2026

Multi-Agent Workflows Use Cheap Proxy Evaluators With Periodic Full-System Tests To Accelerate Kernel Evaluation

Advancing science & engineering with Antigravity’s /teamwork · The latest research from Google

Science, Technology & Innovation · Nov 13, 2026

Multi-agent workflows are producing measurable infrastructure gains: validated 23-second TPU kernel proxies enabled faster optimization and yielded 4.09× speedups on JAXBench and 2.99× on vLLM Pallas kernels, while Borg scheduling reduced TPU preemption by 24% without harming SLOs or utilization.


11/13/2026

Machine-Checked Formalizations Provide An Audit Layer For Long Computer-Assisted Proofs

Advancing science & engineering with Antigravity’s /teamwork · The latest research from Google

Science, Technology & Innovation · Nov 13, 2026

Google’s LEAP multi-agent system reportedly formalized major mathematical results in Lean at million-line scale, showing that theorem specification and proof-pipeline management—not just proof invention—are becoming the main bottlenecks, while machine-checked artifacts and lighter review workflows can expose errors missed by conventional peer review.


11/13/2026

Auditable Candidate-Generation And Physics-Verification Loop Improves Molecular-Structure Identification Over Standalone Prediction

Advancing science & engineering with Antigravity’s /teamwork · The latest research from Google

Science, Technology & Innovation · Nov 13, 2026

Google’s multi-agent system for identifying molecular structures from NMR data separates candidate generation from physics-based verification, achieving 95% top-1 and 100% top-3 accuracy on benchmark molecules—though without laboratory or clinical validation.


11/13/2026

Independent Verification And Task-Boundary Design Drive Reliability In Multi-Agent Systems

Advancing science & engineering with Antigravity’s /teamwork · The latest research from Google

Science, Technology & Innovation · Nov 13, 2026

Google reports that reliable multi-agent research comes from parallel exploration paired with independent executable verification and tightly scoped task handoffs—not from a single model run or longer context—and demonstrated this by formalizing the Courtade–Kumar conjecture in Lean 4, though the individual workflow contributions were not independently isolated.