Google Cloud’s AI agents now work autonomously with voice and guardrails

From chatbots to autonomous AI agents

Rajesh Shewani, who leads AI customer engineering for digital natives, startups, and ITeS at Google Cloud, kicked off his TechSparks 2026 masterclass by tracking generative AI’s evolution. It began with simple retrieval, moved to agents, and has now entered what Google calls the “harness era.” Harnesses wrap AI models in memory, context, and guardrails—letting them run tasks without constant human input. That turns AI from a disappearing assistant into a persistent colleague.

The four building blocks of AI agents

For an AI agent to work, Shewani said it needs four things: a clear goal (book a ticket, update a CRM), reusable instructions, access to company knowledge, and tools to act—like making calls or updating databases. Google is rolling out the Open Knowledge Format (OKF) to standardize that knowledge. OKF stores data as markdown files with YAML frontmatter, making it easier to plug into different systems.

Cost control is critical. Google uses a shared knowledge layer that mixes structured context with plain text, cutting down on expensive token use. It’s the same logic engineers have always used—optimize what you spend to get the job done.

Voice takes over as the new AI interface

Voice is becoming the dominant way people interact with AI, especially in markets where English isn’t the first language. Google’s Bengaluru team has added more than 90 Indian languages and dialects to Gemini Live, including Sanskrit, Bhojpuri, and Maithili.

Shewani showed a live demo built for Grab, Southeast Asia’s largest ride-hailing app. Grab’s drivers and riders already make over 10 million voice calls every month. The demo used Gemini 3.5 Live Translate to convert conversations in real time across more than 70 languages. Back in India, another demo turned a news video into Marathi and Malayalam with a single prompt—syncing lip movements and facial expressions. That slashes the steps needed for dubbing, which once required separate training and marketing pipelines.

Enterprise tools and the rise of AI-generated code

Shewani walked through Google’s Gemini Agent, launched in October 2026. It runs in the cloud for hours or even days, even if the user’s device goes offline. One example: an agent that researches procurement options, drafts a request for proposal, sends it to vendors, evaluates the responses, and delivers a shortlist—switching between Gemini and Anthropic’s Claude depending on the task.

Enterprises want to adopt AI without losing control. Google’s answer is Gemini Enterprise, which builds in guardrails and usage limits. Nearly 90% of Fortune 100 companies now use it, Google Cloud CEO Thomas Kurian told the audience.

Coding is another frontier. Google’s Antigravity harness writes and reviews code, and 75% of the company’s new code is now AI-generated—though engineers still review it. Tools like CodeMender go further: they scan for vulnerabilities, generate patches, test them, and push updates straight to GitHub repositories.

How this fits with earlier AI trends

The shift toward AI-driven development didn’t happen overnight. Earlier stories on Anthropic’s rapid GPU porting and cloud providers’ differing AI strategies foreshadowed today’s focus on guardrails and token costs. As AI infrastructure grows, electricity bills climb, making cost control a top concern for enterprises.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

Google’s push toward autonomous AI agents means startups must design systems that can run without constant oversight—especially if they target voice or multimodal use cases. The cost of tokens and guardrails will decide whether these tools scale or stall.

Key takeaways

• Autonomous AI agents now perform end-to-end workflows, from research to execution, without human intervention.
• Voice and multimodal tools are becoming essential, with Gemini Live supporting over 90 Indian languages to reach non-English speakers.
• Nearly 90% of Fortune 100 companies use Gemini Enterprise, signaling that cost control and security are make-or-break for AI adoption.
• Startups must adopt knowledge standards like OKF and optimize token use to avoid spiraling costs as AI agents handle more tasks.

Related on Lazyfounder

• Google DeepMind
• Google
• Anthropic
• More ai news
• Anthropic claims AI-driven tools slash AMD GPU porting to three days
• AWS, Microsoft, and Google Cloud reveal sharply different AI bets in Q3