Google describes two systems projects where people set architecture and acceptance criteria while agents implement and verify: FarArray for distributed graph computation and a TPU sparse graph kernel that avoids edge-sized intermediates.
Source: The Antigravity & Google Research Team, How Gemini 4 Argon Upleveled Agents on Antigravity: Trillion-Edge Graphs and Custom TPU Kernels, October 26, 2026.
FarArray exposes pinned cluster DRAM as a partitioned global address space, using RDMA for direct remote reads, writes and 64-bit atomics to reduce remote CPU and general-purpose key-value overhead. Thread-local pruning prevents most redundant edges from reaching the network. On two connected-components benchmarks Google reports 40–180 times faster runtime and 6–22 times less RAM than TCP key-value baselines. A separate 1.14-trillion-unique-edge case uses 148 GB of pinned memory across 64 targets. These tasks and baselines should not be collapsed into one speedup.
The Pallas TPU kernel streams COO edges in blocks, avoiding the edge-count-by-feature-dimension HBM intermediate produced by a conventional gather and segment sum; the backward pass retains integer edge indices. The authors report 8–15 times faster forward and backward passes on TPU v6e, with temporary HBM falling from 20.2 GiB to 111.5 MiB for 20 million edges, and claim numerical parity.
The team says humans defined interfaces, memory layout and no-mock verification requirements, while proposer and verifier agents implemented, debugged and checked the work. This is an internal case study with company-reported benchmarks; the claim that months of work were compressed into hours depends on project estimates and does not establish a general development speedup.