The Agentic OS

Demonstrating visionary, high-level strategic systems thinking and cross-functional leadership, this project defines the future of AI agents on Android. By designing the AI Co-Operating System, we unified disparate AI agents into a cohesive, privacy-first framework that transforms Android from a reactive tool into an intelligent, proactive co-pilot.

The Challenge of Fragmentation: The "Layered Mess"

When large language models and context-aware systems proliferated, every functional team tried to claim a visual anchor on Android. Voice overlays, corner gestures, long-press power button configurations, keyboard suggestion rows, and home screen widgets began vied for screen real estate.

The immediate result was a chaotic "layered mess" of overlays that competed for user attention, triggered conflicting automation routines, and caused severe organizational friction regarding spatial anchors. Instead of an elegant, intelligent assistant, the system began to feel cluttered, overwhelming the user with decision fatigue and choice overload.

The Process: Spatial & Platform Reconciliation

To break the strategic deadlock and align our roadmap, I designed and led cross-Product Area (cross-PA) workshops. This initiative gathered designers, engineers, product managers, and researchers from Android SysUI, Privacy UX, Assistant/Gemini, and external design agencies.

By mapping the entire mobile interface lifecycle—from static screen states to active voice triggers and multi-step background tasks—we established shared spatial principles. We formulated platform guidelines to treat the operating system not as a launchpad for isolated features, but as an integrated visual mediator.

The Solution: A Unified AI OS Framework

The result of this effort is a pioneering framework that mediates context sharing across the platform. The architecture centers around two primary systems thinking components:

1. Distributed Attention Layers

To resolve spatial clutter, we structured all intelligence features into three strict, hierarchical attention layers:

• Ambient: Passive, background intelligence processing signals (such as device state, location, and physical environment) to surface micro-actions without breaking flow.
• Contextual: On-screen viewport understanding that parses active content locally through Private Compute Core, surfacing relevant next steps dynamically.
• Agentic: Active, multi-step orchestration where the system operates across multiple apps on the user’s behalf, keeping the user in full control through continuous observability.

2. The Spectrum of Automation

Rather than treating automation as a binary state, we defined a spectrum of agentic actions. This provides designers and engineers with strict guidance on where and how to assert intelligence:

• Enhance: Elevate user actions (such as dictation, writing styles, and system inputs) while keeping the control loop fully manual.
• Automate: Take over highly predictable, repetitive tasks (like form completions or meeting logistics) where risk is minimal.
• Recommend: Proactively surface potential transactional opportunities based on ambient data, asking the user to confirm execution.
• Step Aside: Intervene and yield absolute control back to the user when facing high-security gates, payment authorizations, or high-ambiguity pathways.

Strategic Platform Impact

The primary value of this framework was establishing a shared conceptual foundation across historically siloed teams. By defining clear boundaries for distributed attention and spatial anchors, the project resolved active alignment deadlocks between System UI, Privacy UX, and third-party developers. Rather than just shipping an isolated feature, this work helped shift Android's underlying model toward a more coordinated, privacy-first system, providing the interaction guidelines and mediation architecture that continue to shape agentic integration on the platform.