Google and Microsoft expand AI agents

From answers to actions

The latest AI race is no longer only about who gives the best answer. It is about who can complete the task, preserve context, ask for permission at the right moment and leave an audit trail afterward. Over the past 72 hours, Google and Microsoft both pushed that idea forward from different starting points: Google through richer conversational interfaces and the agentification of its largest distribution surfaces, Microsoft through a redesigned Copilot aimed at becoming an operational layer for work [1][2][3].

Google’s move has two visible parts. First, it introduced Gemini 3.8 Live with Live Avatar, a feature that gives enterprise conversational AI a real-time visual persona able to listen, see and speak with expressive audio and video [1]. Second, reporting on Google’s broader product shift says the company has turned its biggest product into an agent platform, a framing that matters because Google’s advantage is not only model quality, but default placement across products people already use [2].

Microsoft’s move is more explicitly workplace-oriented. The company introduced a new Copilot built around Home, Code and Autopilot, with Office brought directly into the Copilot experience and long-running agents designed to perform delegated tasks rather than simply draft replies or summarize documents [3]. Taken together, the two announcements point to the same strategic destination: agents that act inside established platforms, not beside them.

Google gives the agent a face

Google’s Live Avatar is not a cosmetic add-on. The company describes it as a near real-time visual presence for Gemini’s live dialogue models, combining speech with streaming video so the agent can appear as a dynamic persona rather than a disembodied voice [1]. The target use cases are practical: customer service, virtual assistance and interactive walkthroughs, where a face, voice and responsive turn-taking may make the interaction feel easier for users who do not want to operate a form, menu or chatbot [1].

The feature also shows how Google is trying to solve a classic interface problem. Voice agents can be fast, but they often feel abstract; video agents can be engaging, but they risk latency, cost and the uncanny valley. Google is betting that precise lip-syncing, natural expressions and fluid turn-taking will make avatar agents useful rather than distracting [1]. That is a product challenge as much as a model challenge: the experience only works if the avatar stays responsive while tool calls, data retrieval and workflow actions happen in the background.

The most important technical detail may be asynchronous tool execution. Google says Live Avatar can trigger tool calls and fetch data while maintaining an active conversation, which is exactly the difference between a demo character and a working support agent [1]. A hotel check-in agent, for example, cannot simply sound friendly; it must verify a reservation, update systems, handle exceptions and keep the customer informed while those actions complete.

Google also built the feature for multilingual scale, saying Live Avatar can transition across 97 languages while adapting lip sync and expressions [1]. That matters for customer support because the economics of AI agents become more attractive when a single system can serve many regions without separate voice, video and language stacks.

Platform power is Google’s real lever

The avatar is the visible hook, but the platform story is the bigger strategic point. Google’s largest advantage is distribution: Search, Gemini, Workspace, Android, Chrome and enterprise cloud products create many places where an agent can appear without asking users to adopt a new destination. The Forkast report framed Google’s shift as turning its biggest product into an agent platform, signaling that the company is moving from answering queries toward completing interactive tasks [2].

That shift is larger than a search-result redesign. Traditional search organizes links and snippets. Agentic search must maintain intent, use tools, ask clarifying questions, act across services and return a completed outcome. If Google can make that feel native inside its existing products, the company gains a powerful adoption path: users do not need to learn “agent software”; they simply encounter more capable behavior inside familiar surfaces.

The risk is that richer interfaces can slow the experience down. A text answer can be almost instant; an agent that sees, speaks, reasons, calls tools and renders an avatar has more points of failure. For Google, the challenge is to prove that interactivity improves outcomes without adding enough latency, cost or awkwardness to make users retreat to simple search. Clippy may indeed want an avatar upgrade, but customers will still judge the agent by whether it solves the problem.

Microsoft turns Copilot into a work operating layer

Microsoft’s announcement is aimed squarely at the enterprise workflow. The new Copilot app introduces Home as the starting point where Chat and Cowork come together, Code as a natural-language way to build apps, automations and workflows, and Autopilot as a persistent agent that keeps working without continuous prompting [3]. Microsoft says Home and Code will roll out through its Frontier program in the coming weeks, while Autopilot is expanding to private preview at the end of September [3].

The product language is revealing. Microsoft is no longer positioning Copilot only as a drafting assistant inside Word or a meeting summarizer inside Teams. It is describing a single place to ask, delegate, build and automate [3]. That moves Copilot closer to an operational layer: a system that can draft a launch brief, update a workbook, build an internal tracker, monitor a channel, follow up on threads and manage recurring work.

Code is especially important because it lowers the boundary between knowledge work and software creation. Microsoft says users can describe an app, tracker, dashboard, automation or workflow in natural language, and Copilot can choose an approach and build it in a sandboxed environment that can be hosted securely inside the customer’s tenant [3]. In enterprise terms, this is where “AI assistant” starts to blur into “internal software factory.”

Autopilot pushes even further. Microsoft describes it as a cloud-hosted digital teammate with its own identity, memory, computer and workspace, visible inside Teams, Outlook, chats, channels and documents [3]. The corporate significance is obvious: if an agent has identity and memory, it can be assigned responsibility. But that also raises the stakes for permissions, governance and cost control.

Governance becomes the adoption test

The more agents can do, the less acceptable it becomes for them to be black boxes. Microsoft’s Copilot announcement repeatedly points to permissions, audit and governance, and its separate run-assert-eval release shows why that matters [3][4]. The new skill is designed to discover agent risks, measure failures, generate runtime policy and rerun evaluations to prove whether the fix worked [4].

That is the kind of machinery enterprises will need before they allow agents to touch sensitive systems. A drafting assistant can be corrected by a human. An agent that changes billing details, contacts suppliers, updates CRM records or pulls customer data needs policy enforcement before and after tool use. Microsoft’s example of a billing-support agent focuses on cross-customer data exposure, a failure mode that would be unacceptable in production [4].

Google faces a similar trust problem, especially as avatar agents enter customer-facing settings. A friendly face can increase engagement, but it can also increase misplaced trust. Google says Live Avatar outputs are watermarked with SynthID to make AI-generated audio and video detectable, which is a necessary safeguard when synthetic faces and voices become part of normal service interactions [1].

The bottom line

Google and Microsoft are expanding AI agents from opposite ends of the same market. Google is emphasizing interface richness and massive distribution: make agents more natural, visual and embedded across high-volume products. Microsoft is emphasizing workplace execution: make Copilot the place where employees ask, delegate, build and automate under enterprise controls.

The winners will not be decided by demos alone. Enterprises and consumers will measure whether agents save time, reduce friction and complete tasks reliably. The hard questions are now operational: Who approved the action? What data did the agent see? What did it change? How much did it cost? Can the company prove the agent behaved within policy?

The answer will determine whether this generation of AI agents becomes a durable platform shift or another layer of productivity software people admire but avoid. For now, the signal is clear: Google and Microsoft are both betting that the next interface is not a search box or a document pane. It is an agent that can see the task, talk through it and get the work done.

Sources from the last 72 hours

1. [1]Introducing Gemini 3.8 Live with Live AvatarSep 24, 2026, 12:00 AM UTC
2. [2]Google Just Turned Its Biggest Product Into an Agent PlatformSep 26, 2026, 12:00 AM UTC
3. [3]Microsoft unveils new Copilot with AI agents, coding tools to expand workplace automationSep 25, 2026, 12:45 AM UTC
4. [4]Introducing the new Copilot with Home, Code and AutopilotSep 25, 2026, 12:00 AM UTC
5. [5]Introducing run-assert-eval: Find the risk, fix it, prove itSep 24, 2026, 12:00 AM UTC
6. [6]Google Launches Conversational AvatarSep 25, 2026, 12:00 AM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.