Google AI Updates: September 26, 2026

1. AlloyDB Spins Up Sandboxed Postgres Instances for Agents on Shared Storage

Google Cloud put PostgreSQL for agents in AlloyDB into preview, an architecture that provisions isolated serverless database instances in seconds so agent queries never touch the production primary, standby, or read replicas. The sandboxes read the same data through Colossus-backed shared storage, and Google claims sub-millisecond I/O, terabit-per-second aggregate scan throughput, and more than 3 million queries per second, scaling to zero when agents finish. Billing is pay-as-you-go for active reasoning loops, and the instances keep full Postgres compatibility including vector, full-text, and spatial search plus federated queries into BigQuery and Spark. Source

2. API Gateway Now Serves Existing REST APIs as MCP Tools

Google Cloud API Gateway added public preview support for the Model Context Protocol, transcoding MCP JSON-RPC calls into REST requests without a separate MCP server. Developers annotate OpenAPI 3.0 or 3.1 specs with x-google-mcp-tool directives, the gateway serves MCP on a /mcp path, and existing auth, quota, and logging policies apply to both traffic types. Limits in the preview include a cap of 1,000 tools per gateway, no support for operations that return HTTP 204, and no streaming responses or MCP resources and prompts yet. Source

3. Google Open-Sourced an Agent That Migrates EKS Workloads to GKE via Pull Requests

Google Cloud released GKE agentic migration, an open-source plugin on GitHub that uses LLM reasoning to translate AWS EKS setups into GKE equivalents, such as IRSA to Workload Identity, ALB to Gateway API, and Karpenter to node auto-provisioning. Generated code passes deterministic checks like terraform validate and manifest contracts before human review, and the tool never touches live clusters, delivering all changes as pull requests. Data movement is deliberately handed off to Database Migration Service and Storage Transfer Service rather than done by the agent. Source

4. MaxText Reproduced OLMo 3 7B Pre-Training on Ironwood TPUs Within 0.012 Loss

Google reproduced AI2’s OLMo 3 7B from scratch in MaxText, covering the roughly 5.93T-token stage-1 pre-training on Ironwood TPUs and the 100B-token mid-training stage on TPU v5p. Training loss stayed within 0.012 of the reference through 800k steps, downstream accuracy stayed within 0.005, and the runs hit 44.5% MFU on Ironwood and 57.4% on v5p. Reshaping attention from 32x128 to 16x256 heads at the same parameter count gave a 12.4% speedup, and the OLMo 3 implementation, conversion tooling, and launch scripts are now in the MaxText repo. Source

5. Google Research Chains Four Systems to Generate 10-Minute Coherent Videos

Google Research described a multi-agent pipeline for long-form video that combines four pieces: an AI Video Co-Director that uses multi-armed bandits to pick creative configurations, CANVAS for persistent memory of characters and object states, A2RD for segment-by-segment autoregressive generation, and VQQA, which uses generated visual questions as feedback to refine prompts. The team reports an 81.4 score on GenAD-Bench for the Co-Director and 10-minute videos that keep characters and environments consistent. The work is published as four arXiv papers, with no code or model release announced. Source