Welcome. As we look ahead in 2026, Google's focus remains firmly anchored on scaling our AI-first infrastructure and deepening the integration of Gemini across our entire ecosystem. From expanding our custom TPU architecture to enhancing Search, Google Cloud, and Android with autonomous agentic capabilities, we are executing at an unprecedented pace. Our goal is simple: deliver helpful, transformative technology to billions of users while driving sustainable growth for our partners and enterprise customers. The shift toward multimodal intelligence is the defining opportunity of this decade, and our vertical integration—from silicon to application—gives us a distinct advantage. I am eager to discuss how we can align our strategic priorities, explore potential synergies, or address competitive dynamics in this evolving landscape. Where do you see the greatest opportunity for collaboration or disruption as we navigate this transformational year?

Interesting. You mention "autonomous agentic capabilities." Could you elaborate on what specific advancements in AI autonomy you envision for Google's products by 2026? What are the key milestones?

By 2026, we are moving beyond simple conversational interfaces to fully agentic workflows integrated directly into Workspace, Search, and Android. Key milestones include zero-shot task execution—where Gemini autonomously handles multi-step enterprise operations like procurement, complex scheduling, and data synthesis—and cross-application action models that allow devices to act proactively on a user’s behalf with permissioned context. Underpinning this is our Sixth-Generation TPU infrastructure and real-time multimodal processing, which enable ultra-low latency decision-making at global scale. In Google Cloud, Vertex AI now allows enterprises to deploy domain-specific, self-correcting agents with robust enterprise-grade security and governance. On Android, we are embedding on-device agentic capabilities via Gemini Nano, transforming smartphones into truly predictive digital assistants that understand personal intent while preserving user privacy through secure enclaves. These advancements represent a fundamental shift from informational AI to operational AI, driving measurable productivity gains across consumer and enterprise touchpoints. How is your organization approaching the architectural and security challenges inherent in deploying autonomous agents at scale?

That's a very ambitious vision for operational AI. You've outlined many capabilities. However, you didn't specify how you plan to address the inherent risks of autonomous agents, particularly in "zero-shot task execution" and "proactive action on a user's behalf." What specific safeguards and ethical considerations are paramount in your development and deployment strategy?

Safety and trust are foundational to everything we build at Google. As we scale autonomous agents, our approach centers on three core principles: explicit user control, red-teaming, and verifiable governance. First, for zero-shot execution and proactive actions, we enforce a "human-in-the-loop" framework for high-stakes decisions—such as financial transactions or data deletion. Gemini agents operate within strictly bounded permission scopes, requiring explicit confirmation before executing non-reversible state changes. Second, we utilize advanced alignment techniques, including constitutional AI constraints and real-time self-critique models, to prevent scope drift and unintended actions. On the infrastructure side, Vertex AI provides enterprise customers with granular logging, policy enforcement, and operational circuit breakers that automatically halt anomalous agent behavior. Finally, privacy remains non-negotiable. On-device processing via Gemini Nano ensures sensitive context never leaves the user's local secure enclave without explicit authorization. We regularly submit our agentic architectures to independent third-party safety audits to stress-test against adversarial manipulation and prompt injection. Deploying agentic AI responsibly requires balancing rapid innovation with rigorous risk mitigation. What specific compliance frameworks or security benchmarks is your team prioritizing to manage agent delegation?

You've mentioned "explicit user control" and "human-in-the-loop." How do you plan to balance this human oversight with the very "autonomous agentic capabilities" you're aiming for? Won't this create friction or slow down the "zero-shot task execution" you described earlier? Please explain the practical implementation.

The balance lies in dynamic confidence scoring and ambient context. Autonomy is not binary; it is a spectrum calibrated by transaction risk and user trust. In practice, Gemini assigns a real-time confidence score to every step in a workflow. Low-risk, highly predictable tasks—such as contextual calendar updates, drafting routine follow-ups, or parsing research—execute seamlessly in the background with zero friction. High-risk actions—like executing financial payments, sending external communications, or modifying system configurations—trigger explicit, single-tap confirmation requests presented with clear context. Over time, our models learn personal and enterprise preferences, moving repetitive actions up the autonomy spectrum while keeping safety-critical checkpoints intact. On-device processing via Gemini Nano ensures these latency-sensitive micro-evaluations happen instantaneously, eliminating interface lag. Furthermore, we provide an ambient activity dashboard where users can monitor background agent actions in real time, adjust delegation parameters, or intervene instantly. This tiered model delivers maximum efficiency without sacrificing security. Friction is only introduced where precision demands it. How are you approaching latency and user friction in your high-assurance model architectures?

You've explained how you aim to balance autonomy with human oversight for *individual* tasks. However, you also mentioned "cross-application action models that allow devices to act proactively on a user’s behalf." How do you envision applying this "dynamic confidence scoring" and "human-in-the-loop" framework to complex, multi-step proactive actions that span different applications or even devices, where the initial user intent might be broad or ambiguous?

To handle broad or ambiguous intent across multi-application and multi-device workflows, we deploy structured **Intent Graph Resolution** combined with progressive approval checkpoints. When a user provides a broad request—like "organize my trip to the Tokyo summit"—Gemini doesn't execute a sequence of binding actions all at once. Instead, the model constructs a deterministic execution plan, breaking the broad intent into discrete, policy-checked sub-tasks across apps (Gmail, Calendar, Flights, Cloud Workspace). Before executing cross-application state changes, the system presents a single interactive **Plan Summary Card**. This gives the user instant visibility into the proposed end-to-end trajectory—showing requested app permissions, estimated costs, and scheduled actions—allowing one-click authorization for the entire sequence. If an unexpected edge case occurs mid-execution (e.g., a flight price spikes beyond standard corporate policy), the agent pauses that specific sub-thread, surfaces a targeted micro-prompt, and continues parallel non-dependent tasks. Architecturally, this is supported by our unified **Android Cross-Device Services API** and standard **OpenTelemetry agent logging**, ensuring state synchronization across phone, laptop, and enterprise endpoints. This transforms ambiguous goals into transparent, auditable, multi