How Gemini 4 Argon Boosted Antigravity Agents: Trillion-Edge Graphs and Custom TPU Kernels

At a glance

Google researchers showed how Gemini 4 Argon and Antigravity's Teamwork, a framework where AI agents split up roles and work together, cut months of low-level networking code close to the hardware and code for TPUs, Google's AI chips, down to hours. Humans set the design and acceptance criteria, while agents handled building and checking the code.

Key points

1. 1

   A distributed systems project expected to take a year was wrapped up in five hours, producing a large-scale graph platform called FarArray.

2. 2

   FarArray handles graphs with more than a trillion edges, the links between points, and delivers 60 to 90 times the throughput of older TCP key-value setups.

3. 3

   Custom TPU kernels, the core math code that runs on the chip, used up to 185 times less memory and ran graph neural network passes 8 to 15 times faster while keeping results identical.

4. 4

   Humans set the design and verification rules, while proposer and adversarial verifier agents took care of coding, debugging, and tuning.

Why it matters

It shows AI agents can turn research ideas that used to wait months or years for rare expertise into working systems much faster.