Google launches Gemini Agent Platform, eighth-generation TPUs

Taming agent sprawl with central governance

For enterprise IT teams, the rapid pace of AI adoption has threatened to create a shadow IT crisis of ungoverned bots and fragmented data silos. Google’s answer is the Gemini Enterprise Agent Platform, which provides a unified framework for developers to build, test and safely deploy agents that can execute multi-step workflows.

The platform includes an Agent Registry for indexing internal tools, and an Agent Gateway that acts as “air traffic control”, providing centralised, real-time policy enforcement and ensuring compliance. To handle complex business logic, the platform supports agentic orchestration, allowing specialised agents to delegate tasks to one another deterministically.

Google is also heavily leaning into open standards, incorporating the model context protocol (MCP) to allow agents to seamlessly interact with Google Cloud services and third-party software-as-a-service (SaaS) applications.

“When we gathered at Next a year ago, we talked about how generative AI was transforming organisations around the world,” said Thomas Kurian, CEO of Google Cloud in a blog post. “Today, that future is in-production – the agentic enterprise is real – and deployed at a scale the world has never before seen.”

To alleviate integration headaches for CIOs, Kurian noted that instead of piecing together a collection of fragmented systems, Google offers a “vertically optimised stack where everything is co-developed”.

Across the Asia-Pacific region, early adopters are already putting agentic workflows into production.

For example, Singapore supermarket chain FairPrice Group has integrated Gemini-powered AI agents into its smart carts to elevate the grocery shopping experience, while Australian financial services firm Macquarie Bank is leveraging Gemini Enterprise to reclaim more than 100,000 hours of team members’ time.

In Japan, gaming giants Square Enix and Capcom are using the Gemini Enterprise Agent Platform to build real-time in-game companions and playtesting agents.

Silicon optimised for training and inference

To support the massive compute requirements of continuously running, autonomous agents, Google expanded its AI Hypercomputer infrastructure with its eighth-generation TPUs, splitting the hardware into two purpose-built chips, TPU 8t and TPU 8i.

Designed for scale, TPU 8t uses breakthrough inter-chip interconnect (ICI) technology to network up to 9,600 TPUs and two petabytes of shared high-bandwidth memory in a single superpod. Google claims it delivers three times the processing power of the previous Ironwood generation.

TPU 8i is built specifically for the low-latency requirements of agentic reasoning, featuring three times more on-chip SRAM to host larger key-value (KV) caches entirely on-silicon. It also uses a new topology to directly connect 1,152 TPUs in a single pod, delivering an 80% improvement in performance per dollar for inference workloads.

Google Cloud also announced it will be among the first providers to offer instances based on the Nvidia Vera Rubin NVL72 systems, alongside its newly launched Axion Arm-based central processing units.

Agentic SecOps and cross-cloud data

Recognising that AI agents are only as effective as the data they access, Google Cloud introduced the Agentic Data Cloud, which includes the new Cross-Cloud Lakehouse based on the open-source Apache Iceberg table format for analytics datasets. The lakehouse allows organisations to run AI queries on data residing in Amazon Web Services – and soon, Microsoft Azure – without the cost and friction of data egress or supplier lock-in.

“You can keep your analytic calculations in one cloud, but connected to data on other clouds, and be able to do so without copying any data from the other clouds,” Kurian noted during a media briefing. “Zero copy for analysis purposes.”

Security, naturally, is a concern as agents are granted access to sensitive corporate data and systems of record. To secure agentic workflows, Google announced Agentic Defense that merges Google’s SecOps and threat intelligence capabilities with Wiz’s cloud and AI security platform.

The integration includes Wiz’s new AI application protection platform, which provides runtime protection for AI workloads. Security teams can leverage specialised Wiz AI agents, including Red Agents to validate exploitable risks, Blue Agents for threat investigations, and Green Agents to automatically generate code-level remediations for vulnerabilities.

Google is already operating at this new agentic scale internally. According to Pichai, the company uses security agents to automatically triage tens of thousands of unstructured threat reports each month, reducing threat mitigation time by more than 90%.

In addition, 75% of all new code at Google is now AI-generated and approved by engineers, up from 50% last fall, driven by fully autonomous digital task forces orchestrating code migrations at unprecedented speeds, he noted.