Reflections on Google Cloud Next ’25

(Source: Google Cloud Next 2025 Wrap Up | Google Cloud Blog)

Disclaimer: Google Cloud provided me with a conference pass and support for travel and accommodation, as with the cases with all invited analysts.

Executive Summary

Google Cloud Next ’25 presented a narrative focused on operationalizing AI through sophisticated “Agents”. This report analyzes the key AI-related announcements from the event, particularly concerning AI Agents and Responsible AI (RAI), evaluating Google’s progress against the assessment from 2024, which characterized its AI story as a “home listed, interiors incomplete”.

While Next ’25 demonstrated substantial advancements in providing the tools and infrastructure for building and deploying AI agents — notably through the introduction of the Agent Development Kit (ADK), the Agent2Agent (A2A) protocol, Vertex AI Agent Engine, and enhancements to Google Agentspace — significant questions persist regarding practical implementation, ecosystem maturity, and pricing transparency. The launch of the Ironwood Tensor Processing Unit (TPU), purpose-built for inference, significantly bolsters Google Cloud’s Generative AI Stack — a key part of what Sundar Pichai called ‘the full stack of AI Innovation’, with homegrown options for every part of that stack.

(Google Cloud AI Stack, Source: Google Cloud Next 25 Opening Keynote)

However, despite filling some gaps identified in 2024, the path to widespread, seamless enterprise adoption of Google’s advanced agentic vision requires further clarification. Key ambiguities around the pricing of agent platforms, the general availability of critical components, the governance and technical specifics of the A2A protocol, and the real-world manageability of complex multi-agent systems remain.

Google Cloud has laid down more sophisticated blueprints and supplied advanced construction materials for its AI “home.” Still, the ease of “moving in” and the final cost of furnishing remain less defined.

Setting the Stage: Revisiting the ‘Incomplete Interiors’

To accurately evaluate the progress unveiled at Google Cloud Next ’25, it is essential to revisit the state of Google Cloud’s AI narrative as assessed last year. Our analysis characterized Google’s position with the analogy of a “home listed, interiors incomplete”. This assessment acknowledged the strength of Google Cloud’s foundational elements — its robust AI infrastructure, comprehensive development tools like Vertex AI, advanced data management capabilities, security frameworks, and collaboration platforms. However, it argued that the crucial layer of ready-to-deploy, enterprise-grade Generative AI offerings was underdeveloped. The core criticisms centered around the lack of out-of-box GenAI offerings/ turnkey solutions and the need for a non-trivial amount of development efforts.

The 2024 assessment concluded that while Google Cloud possessed immense potential underpinned by robust technology, it needed to bridge the gap between its platform capabilities and the immediate requirements of enterprises seeking tangible AI-driven outcomes. The foundation was solid, but the readily habitable “rooms” were few. This context provides the baseline against which the announcements and strategic direction presented at Google Cloud Next ’25 must be measured.

Google Cloud Next ’25: The Agentic Era Takes Command

Google Cloud Next ’25 unequivocally positioned AI Agents not merely as a feature but as the central paradigm for the future of enterprise AI interaction on its platform. The event’s keynotes and myriad announcements painted a picture of an impending “Agentic Era,” moving beyond reactive AI assistants towards proactive reasoning systems capable of complex task execution and collaboration.

Overview: Agents Take Center Stage

The narrative shift was palpable. Google Cloud executives, including CEO Thomas Kurian, consistently emphasized the development and deployment of intelligent agents designed to understand goals, plan, remember context, and utilize tools autonomously. This vision was explicitly linked to the Agentic Era where AI systems/agents transition from simply responding to prompts to proactively generating insights and interpreting data, often collaboratively.

This focus wasn’t purely theoretical; Google Cloud highlighted numerous real-world examples of agents delivering business value. These included customer service agents handling large volumes of interactions for companies like Wendy’s (60,000 drive-through orders daily) and Lowe’s, navigation agents in Mercedes-Benz vehicles, expert guidance agents like The Home Depot’s “Magic Apron,” data agents automating tasks at Wayfair and AES, and security agents monitoring threats for firms like Dun & Bradstreet (check out the link above for more use cases demonstrated).

The overarching message was clear: Google Cloud is investing heavily in enabling sophisticated, multi-agent systems capable of tackling complex enterprise challenges, aiming to embed this agentic capability across its platform and into the daily workflows of employees.

Deep Dive: Vertex AI Agent Builder — The Developer’s Toolkit

At the heart of Google’s strategy for enabling agent development lies the significantly enhanced Vertex AI Agent Builder platform. Positioned as the core toolkit for developers, it comprises several key components designed to streamline the creation, deployment, and management of AI agents:

  • Agent Development Kit (ADK): Introduced as a cornerstone announcement, ADK is an open-source framework aimed at simplifying the construction of both single agents and complex multi-agent systems. Google Cloud claims developers can build functional agents in under 100 lines of intuitive Python code (with support for other languages planned). ADK provides developers with deterministic guardrails and orchestration controls for precise management of agent behavior, including how they think, reason, and collaborate. It also supports unique bidirectional audio and video streaming capabilities for more human-like interactions. ADK is also designed to be flexible, allowing integration with popular open-source frameworks such as LangChain, LangGraph, AG2, and Crew.ai, enabling developers to leverage familiar tools within the Google Cloud ecosystem. The developer keynote prominently featured ADK in a demonstration building a kitchen remodeling agent that generated a complete proposal for remodeling on PDF based on the specs improved using Gemini.

(Google Cloud AI Stack, Source: Google Cloud Next 25 Developer Keynote)

  • Agent Engine: Complementing ADK is the Agent Engine, a fully managed runtime environment within Vertex AI specifically designed for deploying agents into production. It abstracts away the complexities of infrastructure management, handling scaling, security patching, monitoring, and reliability automatically. Agent Engine supports agents built with ADK or other frameworks and is designed to maintain conversation context through both short-term and long-term memory capabilities. The developer keynote showcased its role in deploying the multi-agent systems.
  • Agent Garden: Developers can leverage resources provided by Agent Garden — which offers a collection of ready-to-use agent samples, pre-built tools, and connectors — to jumpstart their projects and learn from established patterns.
  • Model Context Protocol (MCP) Support: ADK’s support for the Model Context Protocol (MCP) was highlighted as a key feature for standardizing how agents interact with models and access necessary data and tools. A2A protocol aims to support a unified structure for information exchange, reducing the need for custom integrations for each tool or data source, a common friction point in multi-agent deployments.

Collectively, these components represent Google’s attempt to create a structured yet adaptable environment for agent development. By offering an open-source ADK, Google Cloud encourages broad adoption and community contribution. Simultaneously, the managed Agent Engine provides the necessary enterprise-grade runtime, addressing operational concerns. The inclusion of support for existing frameworks like LangChain is a pragmatic acknowledgment of the current developer landscape, lowering the barrier to entry and positioning Agent Engine as a viable deployment target even for agents not initially built with ADK. This layered approach seeks to balance developer flexibility with enterprise requirements for robust, manageable deployments.

Fostering an Ecosystem: The A2A Protocol

Recognizing that enterprise AI environments will likely involve agents from multiple vendors and built on different frameworks, Google Cloud introduced the Agent2Agent (A2A) protocol. Positioned as an open standard initiated by Google Cloud — claimed as a first among hyperscalers — A2A aims to establish a common language for AI agents to communicate, exchange information securely, and coordinate actions across diverse enterprise platforms. The stated goal is to foster a future of seamless AI agent interoperability, breaking down potential silos and enabling more powerful, automated workflows that span multiple systems and agent types.

The introduction of A2A represents a significant strategic move and invited some A2A vs MCP debates, too. Sundar’s tweet-poll may also have contributed to this debate. The discussion turned out to be much ado about nothing, as A2A works in tandem with MCP and is not meant to be a replacement.

By proposing a standard for agent communication in this nascent field and securing early buy-in from numerous partners, Google Cloud aims to shape the market’s direction. If A2A achieves widespread adoption, it could become the de facto standard, potentially positioning Google Cloud, where A2A is natively supported, as a central hub for orchestrating complex, multi-vendor agent ecosystems. This could create a powerful network effect benefiting Google, even with the protocol itself being open. However, the ultimate success and practical impact of A2A remain contingent on factors yet to be fully clarified, such as the governance model for the standard, and the commitment of vendors to implement it faithfully. Its ability to enable truly seamless and robust cross-platform agent collaboration is yet to be proven in real-world deployments.

Bringing Agents to the Enterprise: Google Agentspace & Workspace Flows

Bridging the gap between sophisticated developer tools and tangible end-user value is critical for enterprise adoption. Google Cloud addressed this “last mile” challenge with significant updates to Google Agentspace and the introduction of Google Workspace Flows (which is in alpha now).

Google Agentspace is positioned as the enterprise platform designed to put Google’s foundation models, search capabilities, AI agents, and actionable knowledge directly into the hands of every employee. Key enhancements announced at Next ’25 include:

  • Chrome Enterprise Integration: Agentspace is integrated directly into the Chrome browser, allowing employees to access enterprise search, data, and potentially agent capabilities within their existing workflow without switching contexts.
  • Agent Gallery: A centralized location for employees to discover and access available agents across the enterprise, whether built by Google, internal teams, or third-party partners.
  • No-Code Agent Designer: A no-code interface for creating or customizing agents to automate specific tasks or enhance knowledge discovery, adapting AI to individual needs.
  • New Pre-built Agents: Google Cloud introduced new first-party agents within Agentspace, such as the “Deep Research” agent (designed to explore complex topics and synthesize reports from internal and external sources) and the “Idea Generation” agent (aimed at autonomously developing and evaluating novel ideas).

Complementing Agentspace, Google Cloud introduced Google Workspace Flows, a new capability to bring agentic AI directly into the familiar Workspace environment (Docs, Sheets, Chat, etc.). Flows are designed to automate multi-step, logic-driven business processes that require context and reasoning, going beyond traditional simple automation triggers. Users can describe the desired workflow in natural language, and Flows can design and build the process, potentially utilizing customized versions of Gemini (referred to as “Gems”) to handle specialized tasks like analyzing marketing copy, reviewing policy documents, or sorting support tickets based on context from Google Drive. Workspace Flows is currently rolling out to alpha customers.

These initiatives directly target the “incomplete interiors” critique from 2024. Agentspace, with its user-friendly interface, pre-built agents, and Chrome integration, provides the accessible front-end that was previously lacking. Workspace Flows embeds agentic automation into the core productivity suite, offering more “move-in-ready” capabilities. However, the ultimate success of Agentspace will depend on its ability to effectively break down enterprise information silos and demonstrate clear ROI beyond enhanced search or basic automation. The effectiveness of grounding mechanisms and the ability of agents like Deep Research to deliver truly insightful synthesis from complex, disparate data sources will be crucial differentiators.

Progress Check: Furnishing the AI Home?

Comparing the announcements at Next ’25 with the specific criticisms leveled in the 2024 “incomplete interiors” assessment reveals significant progress, although questions about full “move-in readiness” persist. It is important to note that most of these capabilities are not generally available yet.

  • Addressing Out-of-the-Box Offerings: The introduction of Google Agentspace, particularly with pre-built agents like Deep Research and Idea Generation, along with Google Workspace Flows embedding agentic automation into productivity tools, represents a clear attempt to provide more readily available solutions. This directly counters the 2024 criticism regarding the lack of such offerings. However, it remains debatable whether the current portfolio of pre-built agents and flows matches the perceived breadth, depth, and seamless integration of competitor offerings like Microsoft’s Copilots, which were highlighted as more “move-in-ready” in the 2024 analysis. Google Cloud appears to be building out its catalog of ready-made solutions, but the ecosystem may still be less mature than established alternatives.
  • Reducing Development Effort: The Agent Development Kit (ADK), with its promise of building agents in under 100 lines of code and integration with Agent Garden’s samples, alongside the no-code Agent Designer in Agentspace, demonstrably lowers the barrier to entry compared to the situation in 2024. These tools offer improved pathways for both developers and business users to create agents. Nonetheless, constructing sophisticated, reliable, and truly autonomous agents capable of handling complex enterprise tasks still necessitates considerable expertise in prompt engineering, tool integration, testing, and managing potential failure modes. While the tools are better, the inherent complexity of building robust agentic systems remains.
  • Moving Beyond the “Framework” Stage: Next ’25 clearly moved beyond merely providing foundational frameworks. While ADK itself is a framework, it is now complemented by Agent Engine, a managed runtime environment for deployment and operation, and Agentspace, a user-facing platform for interaction and discovery. The A2A protocol further aims to provide an interoperability layer. This constitutes a much more complete ecosystem than what was available in 2024. Google Cloud is no longer just offering blueprints and raw materials; it’s providing the construction machinery (Agent Engine) and a showroom (Agentspace).

In summary, Google Cloud has made substantial strides in furnishing its AI “home.” The tools and platforms needed to build, deploy, and interact with agents are significantly more developed and accessible than they were in 2024. The focus remains strongly on enabling customers and partners to build their own solutions, leveraging the flexibility of the platform, rather than Google Cloud itself providing a vast suite of finished, off-the-shelf agents. This strategy plays to Google Cloud’s strengths as a platform provider, but the immediate availability of diverse, proven agent solutions is still heavily reliant on the speed and success of ecosystem development. The “interiors” are being actively furnished, but the selection of readily available, fully decorated “rooms” is still expanding.

Navigating the Details: Pricing, Availability, and Lingering Questions

Despite the ambitious vision and technological advancements showcased at Next ’25, several practical questions regarding pricing, availability, and implementation details remain, potentially impacting enterprise adoption roadmaps.

  • Pricing Clarity Still Evolving:
    Vertex AI Agent Builder discloses base rates — $0.00994 per vCPU-hour and $0.0105 per GiB-hour for memory. Model usage follows Vertex Model Garden’s published token pricing. However, additional costs for integrated tools like Code Interpreter and BigQuery connectors remain unspecified. Google Agentspace pricing is public — NotebookLM Plus Enterprise: ($9/user/month), Agentspace Enterprise ($25/user/month), Agentspace Enterprise Plus ($45/user/month). In contrast, Workspace Flows has no published pricing, suggesting it’s still in experimental phases or subject to custom enterprise deals.
  • Key Features Still in Preview or Alpha:
     Several flagship tools remain unavailable for general use — Ironwood TPU **(**GA expected later in 2025), Gemini 2.5 Pro (in preview on Vertex AI), Gemini 2.5 Flash (in preview on Vertex AI and AI Studio), and most of the innovations related to AI Agents are in various stages of early availability.

Together, these factors signal a robust strategic direction, but the capabilities are still maturing — posing challenges for enterprise teams looking to make immediate, large-scale deployments.

Gaps and Unclear Areas: Beyond pricing and GA dates, several strategic and technical questions linger:

  • A2A Protocol Governance: How will the A2A standard be governed to ensure true interoperability and prevent it from becoming a de facto Google-controlled standard? What are the detailed technical specifications that guarantee seamless communication versus requiring vendor-specific adaptations?
  • Multi-Agent System Management: While Gemini Cloud Assist Investigations was demonstrated for debugging, how mature, scalable, and user-friendly are the tools for managing the lifecycle (monitoring, updating, versioning, performance tuning) of complex, interconnected multi-agent systems in production environments? Real-world complexity often surpasses controlled demos.

These ambiguities highlight that while Google Cloud Next ’25 provided a much clearer strategic roadmap for its AI ambitions, the journey towards making these advanced agentic systems easily adoptable, predictably priced, and demonstrably manageable for the average enterprise is still ongoing.

CloudDon Take

Recommendations for Enterprises

Enterprises evaluating Google Cloud as the foundation for their AI-driven transformations, particularly focusing on the agentic capabilities showcased at Next ’25, should consider the following recommendations:

  1. Define Clear Business Objectives: Before diving into specific technologies like ADK or Agentspace, clearly articulate the business problems you aim to solve or the opportunities you want to capture with AI agents. Focus on tangible outcomes and ROI rather than adopting technology for its own sake.
  2. Assess Platform Maturity and Timelines: Carefully evaluate the General Availability (GA) status of the specific Google Cloud components required for your use case (e.g., Vertex AI Agent Builder components, specific Gemini models, Agentspace features, Workspace Flows). Factor the risks and potential instability associated with relying on features currently in preview or alpha stages into your project planning.
  3. Budget for Uncertainty: Given the current lack of transparent pricing for end-user platforms like Agentspace and Workspace Flows, and potential fees for tools within Agent Builder, build flexibility into your AI initiative budgets. Engage with Google Cloud sales for clarity on preview offerings and possible future costs.
  4. Invest in Skills and Expertise: Recognize that despite advancements like ADK simplifying initial development, building, deploying, and managing sophisticated, reliable, and secure AI agents at scale still requires significant expertise in AI/ML engineering, MLOps, prompt engineering, domain-specific integration, and more. Plan for internal upskilling, targeted hiring, or partnering with experienced service providers.
  5. Start Small, Pilot, and Iterate: Begin with well-defined pilot projects targeting high-impact, lower-complexity use cases. Leverage tools like Vertex AI Agent Builder, the Agent Garden samples, and potentially Google’s pre-built agents to gain practical experience, demonstrate value quickly, and refine your approach before committing to large-scale, complex deployments.
  6. Explore the Ecosystem: Don’t limit your exploration to Google Cloud’s first-party tools. Investigate solutions offered by Google Cloud partners and leverage resources like the Agent Garden and open-source frameworks supported by ADK to accelerate development and potentially find specialized capabilities. Keep a close eye on the development and adoption of the A2A protocol for potential future multi-vendor interoperability.
  7. Prioritize Responsible AI from Day One: Implement robust AI governance frameworks and Responsible AI practices from the outset. Leverage Google Cloud’s grounding capabilities but supplement them with clear internal policies, ethical guidelines, human oversight protocols, and mechanisms for transparency and bias detection. Don’t treat RAI as an afterthought.
  8. Foster Collaboration with Google Cloud: Engage actively with Google Cloud support, technical account managers, and partners, especially when navigating preview features, complex deployments, or unclear pricing structures. Provide feedback to help shape product development.
  9. Stay Agile and Informed: The AI Agents landscape and Google Cloud’s offerings are evolving rapidly. Continuously monitor announcements regarding GA timelines, pricing updates, A2A protocol specifications, new agent capabilities, and RAI tooling to adapt your strategy accordingly.

Strategic Recommendations for Google Cloud

Google Cloud Next ’25 represented a significant evolution in Google’s AI narrative, moving decisively to address the “incomplete interiors” critique from 2024. The event showcased a strong focus on enabling an “Agentic Era,” providing developers with more sophisticated tools (ADK, Agent Engine, Agent Garden) and enterprises with platforms (Agentspace, Workspace Flows) designed to bring AI agents into mainstream workflows. The launch of the inference-focused Ironwood TPU substantially strengthens the foundation of Google’s AI Hypercomputer and its claim to a vertically integrated, homegrown GenAI stack. The proposed A2A protocol signals an ambition to shape the future of agent interoperability.

However, the transition from technological capability to widespread, practical enterprise adoption faces hurdles despite this progress. Key components remain in preview or alpha stages, comprehensive pricing details are lacking (especially for user-facing platforms), the practicalities of governing the A2A ecosystem are undefined, and the real-world manageability of complex multi-agent systems needs further validation beyond initial demonstrations.

Google Cloud’s quintessential challenge has been proving enterprise readiness despite demonstrating cutting-edge technology — which it clearly possesses. This involves ensuring ease of adoption, predictable cost structures, robust manageability, and demonstrable value for the sophisticated agentic systems it envisions. To accelerate the realization of its AI vision and solidify its competitive position, Google Cloud should consider the following strategic recommendations:

  1. Accelerate GA and Clarify Pricing: Prioritize moving key components and flagship models to General Availability, and publish clear, comprehensive, and predictable pricing models for all components — especially the user-facing platforms like Agentspace and Workspace Flows, as well as the associated tools and pre-built agents within Agent Builder. This transparency is essential for enterprise budget planning and adoption decisions.
  2. Flesh out A2A Protocol Details: Move quickly to publish detailed technical specifications for the A2A protocol. Establish and communicate a clear governance model that fosters trust and ensures genuine interoperability, addressing concerns about potential vendor lock-in. Showcase concrete, compelling use cases demonstrating successful cross-vendor agent collaboration enabled by A2A to build momentum and encourage broad ecosystem adoption beyond the initial signatories.
  3. Showcase Multi-Agent Management at Scale: Invest further in developing and prominently demonstrating robust, user-friendly tools and established best practices for the end-to-end lifecycle management of complex multi-agent systems deployed in production. This includes advanced debugging, monitoring, performance optimization, version control, and security management capabilities that address real-world operational challenges at scale.
  4. Mature Out-of-the-Box Agent Solutions: While continuing to empower ecosystem partners, Google Cloud should accelerate the development and release of its own sophisticated, pre-built agents within Agentspace and potentially Workspace Flows. Target common, high-value enterprise use cases across different industries to provide immediate, demonstrable value and lower the initial barrier to entry for customers, effectively complementing the “build-your-own” toolkit approach.

Conclusion

Google Cloud Next ’25 decisively advanced its AI narrative, centering on an “Agentic Era” enabled by tools like the Agent Development Kit (ADK) and a completed homegrown GenAI stack featuring the inference-focused Ironwood TPU. This addressed key gaps identified previously, furnishing the AI “home” more completely than in 2024. However, questions surrounding pricing clarity, component availability, and the practical complexities of managing multi-agent systems indicate that while the technological foundation is robust, the path to effortless, widespread enterprise adoption requires further paving.

Google Cloud has shown the blueprint and supplied advanced materials; now the focus must shift to ensuring a smooth “move-in” for businesses seeking to leverage these powerful new capabilities.