Did someone say AI agents, Google asks, bursting in

Customers aren't sure, economy isn't great, tech looks cute, though

Cloud Next This week Google joined a throng of tech vendors pushing the concept of "agentic AI" on an unsuspecting and perhaps unreceptive collection of enterprise users. Questions remain about how effective this tranche of tools will be at solving business problems and how much it might all cost.

At its annual Google Cloud Next bash in Las Vegas this week, the Chocolate Factory opened the floodgates to a slew of product news.

Among the announcements was a promise to introduce an Agent Development Kit (ADK), an open-source framework Google said would simplify the process of building business software that integrates AI agents – question-answering and task-performing software that makes decisions and forms its output using large language models. Agents can be seen as software that talks to other software as well as people.

Google claims its ADK will allow customers to build an AI agent in under 100 lines of hopefully intuitive code. The ADK also comes with an "agent garden" of pre-built bots and tools, as well as more than a hundred pre-built connectors to common data sources.

Meanwhile, the search and ads giant is set to introduce a new protocol designed to help agents from different vendors talk to one another. The Agent2Agent (A2A) protocol has attracted 50 partners signing up, such as Accenture, Box, Deloitte, Salesforce, SAP, ServiceNow, and TCS, which it said were actively contributing to it.

This is giving enterprises false hope in accelerating human replacement, and we are not there yet

Digging into the details, Google's agent garden is a pre-packaged collection of tools within the ADK allows users to access 100+ pre-built connectors, custom APIs, integration workflows, or data stored within Google cloud systems like BigQuery and AlloyDB. Google cited law firm Freshfields as an adopter of its agent vision. The smaller biz said it would use Google's enterprise data platform for AI, Vertex AI, to create bespoke AI agents for its legal and business processes, while also using search-bots to bring together company-held information.

To power language processing and information gathering for users, Google announced the seventh generation of its Tensor Processing Unit, aka TPU, specialist AI-accelerating matrix-math hardware first used internally at Google.

The chips code-named Ironwood – which you can read more about on The Reg here and The Next Platform here – have more than 10 times the performance of the earlier Trillium TPU. A fully decked out pod of Ironwood TPUs, rented from Google Cloud, can hit 42.5 exaFLOPS of FP8 compute, "meeting the exponentially growing demands" of generative models such as Google's recently announced Gemini 2.5.

Google is not alone in seeing in AI agents an opportunity to extract more revenue from customers. Salesforce, for example, has been extremely bullish on the idea of business interactions being powered by its vision of gen-AI-powered automatons. CEO Marc Benioff told investors last year the company might overcome a fall in customer license numbers by charging users for each and every conversation they have with a bot, which he said was “a very high margin opportunity” for Salesforce.

SaaS business application vendor Workday has also pushed AI agents on its new platform, even seeing them as a means to cut its own human headcount.

Meanwhile, giant omni-vendor Microsoft is also in the agent game with its 365 Copilot range, powered by its $10-billion-plus alliance with ChatGPT maker OpenAI.

But what about users, what are they getting out of it? In a recent research note, Gartner said folks should negotiate Microsoft Copilot Studio products with care as there is potential for extra costs to arise on top of direct licensing.

“Microsoft 365 Copilot Chat could have a significant impact on the software and cloud spend of Microsoft clients. Governance of usage and cost is required,” it said.

Across all AI agent deplyments, users needed to be aware of the complexity of demands they could make on their cloud infrastructure.

Yrieix Garnier, products veep at monitoring and observability vendor Datadog, said agents often operate across multiple platforms, such as Microsoft and Google. "Their behavior can vary significantly based on input, context, and chaining logic," he said. "A single user prompt can initiate a multi-step reasoning process, trigger API calls, or spawn other agents — making usage and cost highly dynamic and difficult to predict.

"To manage this complexity, organizations need to implement observability practices that go beyond traditional monitoring. This includes capturing and analyzing inputs, intermediate reasoning steps, and outputs for every agentic task.

"Token usage, latency, model responses, and error rates must be tracked at a granular level. Just as with microservices, teams need traceability and context to understand where inefficiencies or errors are occurring — and what they're costing. Guardrails are also critical: Without limits, agents can enter loops or generate excessive downstream tasks, leading to uncontrolled spend."

Google's agent interface protocol is quite compelling ... Chances are that this is more marketing hype

Chirag Dekate, VP analyst at Gartner, said Google was differentiating itself in the market with its investment in TPUs, its ability to interact with data sources, and the A2A protocol, which promises collaboration across agents from different vendors.

"If you're trying to create an agented enterprise of the future, you need to have agents working with other agents to solve complex tasks and complex activities. Here, Google's agent interface protocol is quite compelling, because it enables you to have multiple agents work collectively or constructively with one another," he said.

Nonetheless, limitations in the current generative-AI models means they are far from the level of intelligence needed to displace a human workforce. "Chances are that this is more marketing hype. This is giving enterprises false hope in accelerating human replacement, and we are not there yet," he said.

At the same time, organizations may be put off investing in the vision of agent AI offered by Google, Salesforce, Microsoft, and others because of the economic uncertainty created by America's global tariff war, which has wiped trillions of dollars off global stock valuations and prompted warnings of worldwide recessions.

"We are starting to see early signals of enterprises shifting towards cost optimization, as opposed to last year's focus around gen-AI. Businesses will continue to invest in gen-AI but cost optimization will likely take a much more serious tone this year round, because everything is changing. AI adoption rates will suffer, but AI innovation will not suffer because the cloud and hyperscalers are not slowing down. They are doubling down on investments," Dekate said. ®