AI enthusiasts rejoice, for Google has released a new open source agent solution on top of updates to its supercomputing platform in Google Cloud.

The Google AI Hypercomputer now includes support for the Nvidia Collective Communications Library (NCCL), open-source machine-learning (ML) frameworks designed for GPU-to-GPU communication primitives. Google’s optimized iteration for Google Cloud’s integrated supercomputing system is known as NCCL/gIB, and is available on A3 Ultra, A4, and A4X virtual machines (VMs).

For AllReduce communication patterns, which are used when multiple GPUs or processes share data with each other in distributed computing, Google claimed NCCL/gIB can be up to 12x faster than upstream NCCL for certain message sizes.​ This was recorded using AllReduce on a 32-node A3 Ultra (H200) cluster with no background traffic.

Using the same cluster under noisy conditions for AllGather patterns, which gather different pieces of data from each compute entity and distribute the complete collection of all pieces to every entity (or worker), NCCL/gIB delivered around 50% higher performance on larger message sizes than upstream NCCL.

NCCL/gIB interacts with ML frameworks and Nvidia GPUs to optimize performance and gather telemetry data. The tool is integrated with various plugins, such as gIB network for improved load balancing on Google's networks, a custom tuner plugin for selecting optimum tuning options on Google Cloud VMs, and the CoMMA profiler plugin, which provides detailed performance metrics and diagnostic data.

“Although it's possible to use the upstream Nvidia Collective Communications Library on Google Cloud VMs without stability problems, NCCL/gIB is better optimized for Google Cloud, and the performance disparity can be very significant for certain communication patterns, even with the same NCCL parameters,” Google explained.

Announcing A2UI

The hyperscaler also introduced a new open-source project, A2UI, an attempt to generate contextually relevant user interfaces (UIs) for AI agents.

Following on from its Agent2Agent (A2A) protocol, which Google donated to the Linux Foundation earlier this year, the endeavour aims to address challenges related to interoperable, cross-platform, and generative or template-driven UI responses from agents.

“If your agent lives inside your application, it can directly manipulate the view layer (e.g., DOM). But in a multi-agent world, the agent doing the work is often remote … It cannot touch your UI directly; it must send messages,” explained Google, adding that rendering UI from remote and untrusted sources relied on the arduous, complex approach of sending HTML or JavaScript, which was then sandboxed within inline frames.

Google’s A2UI team, therefore, sought a safe way to transmit UI like data but in an expressive code-like format, while matching an app’s native styling.

A2UI thus offers a standard format which can be “generated on the fly” as either structured output or a template hydrated with values.

“The agent generating this response might be a remote A2A agent or the orchestrator the user is interacting with. The JSON payload can be sent to the client over A2A, AG UI, and potentially other transports. The client application renders using its own native UI components ... helping to ensure the agent's output always feels native to your app,” explained Google developers.

A declarative data format ensures a safety-first approach, while also siloing the UI structure from the UI implementation.

The UI is represented as a flat list of components with ID references, making it easier for large language models (LLMs) to output incrementally, supporting progressive rendering and enhancing responsiveness in the user experience.

Google’s A2UI currently works with A2A Extension, AG UI Integration, client libraries such as Web Components, the CoPilotKit agent UI toolkit, and any agent with A2A via the A2A Extension tool.

With the new release, A2UI joins other open source initiatives aiming to standardize the use of the many AI agents out there, with use cases including network and telecom workloads.

These include model context protocol (MCP), which recently saw expanded support within Google Cloud, as well as agentic directory Agntcy, as originated by Cisco, and the Agentic AI Foundation (AAIF), the Linux Foundation’s recently introduced neutral AI development infrastructure.