Inside Google COSMO — The New Architecture of On-Device AI Agents

Inside Google COSMO — The New Architecture of On-Device AI Agents

On May 1, 2026, Google accidentally published a 1.13 GB experimental app called COSMO to the Play Store, then quickly pulled it. Two weeks before Google I/O, this leak revealed the hybrid architecture of next-generation AI assistants.

What Happened

In the early hours of May 1, 2026, Google Research's official Play Store account published a package named com.google.research.air.cosmo. The 1.13 GB download size stood out, and the description read "experimental AI assistant."

Within hours, Google quietly removed the listing — but not before it was downloaded and reverse-analyzed by journalists. The package path research.air reveals it as a next-generation assistant prototype from Google Research's AIR (AI Research) group.

With Google I/O 2026 scheduled for mid-May, this looks like an accidental two-week-early reveal of a planned announcement.

The Core Insight — Three Fulfillment Models

The first thing that catches your eye in COSMO's settings is a user-selectable three-mode processing system:

What "PI" stands for hasn't been officially defined. Personal Intelligence is the most likely interpretation, referring to server-side Gemini models (likely Gemini 2.x Pro or dedicated infrastructure).

What makes this 3-mode design interesting isn't just "strong vs weak model" — it's that the trade-off itself is exposed to the user.

PI Only    → Quality ↑↑, Privacy ↓,   network-dependent
Nano Only  → Quality ↓,  Privacy ↑↑, offline OK
Hybrid     → Quality ↑,  Privacy ~,   auto fallback

Why This Architecture — Comparing to Apple Intelligence

This pattern isn't entirely new. Apple Intelligence uses similar routing:

• On-device: Small tasks run locally
• Private Cloud Compute: Larger tasks go to privacy-preserving servers
• ChatGPT delegation: Complex queries can be routed externally

But while Apple decides routing automatically, COSMO lets users choose. This is the key philosophical difference.

Google's approach is developer/power-user friendly: it shows you which data goes to the server, then lets you choose. This aligns well with EU GDPR and tightening global privacy regulations.

14 Skills — Is This Really an "Agent"?

What separates COSMO from a basic chatbot is its 14 pre-defined Skills, which trigger proactively based on user activity.

Productivity Skills

• List Tracker — track to-do lists
• Document Writer — auto-generate documents
• Calendar Event Suggester — propose schedule entries
• Add Timer — set timers

Research & Knowledge Skills

• Deep Research — in-depth research (likely Gemini's Deep Research feature)
• Google it — delegate to search
• Jargon Definitions — auto-explain technical terms
• Provide Insight — context analysis with insights

Memory & Context Skills

• Recall — recall past conversations/activity
• Conversation Summary — summarize discussions
• People Understanding — learn frequently contacted people
• Event Understanding — learn user's event patterns

Visual & Browser Skills

• Quick Photo Lookup — gallery search delegation
• Browser Agent — web automation (uses Project Mariner)

The fact that Browser Agent uses Project Mariner is decisive. Mariner is Google's browser automation agent, announced in December 2024, which manipulates websites directly through a Chrome extension. COSMO integrates this as one tool inside the on-device agent.

Technical Implementation — AccessibilityService

COSMO uses Android's AccessibilityService API. Originally designed for screen readers and accessibility features, this API is increasingly used by AI agents for screen perception + manipulation.

[User screen] 
    ↓ (AccessibilityService captures)
[Screen context as text]
    ↓
[Skill trigger decision]  ← Nano (cheap) or PI (accurate)
    ↓ (activate appropriate Skill)
[Skill executes → UI manipulation or response]

This is conceptually similar to Apple's App Intents, but applies to all apps, making it more universal. The downside is security — users must grant full screen access, so trust is essential.

What "AIR" Tells Us

The package path com.google.research.air.cosmo contains an interesting clue: AIR. The most likely meaning:

• Agentic Intelligence Research
• Or a codename for a specific Google Research group

This aligns with Google's Agentic AI emphasis throughout 2025 — signaling that the next-gen AI direction isn't just chatbots, but agents that act on the user's behalf.

Developer Angle — The Rise of Gemini Nano API

Here's where it gets really interesting: the architecture COSMO demonstrates is something developers can build too.

Google's Gemini Nano API provides:

• AICore system service — system-level Nano access on Pixel 8+
• ML Kit GenAI APIs — higher-level abstractions
• Summarization, extraction, rewriting, translation as basic tasks

So anyone can write code like this:

// On-device summarization with Gemini Nano
val summarizer = Summarization.getClient(featureOptions)
val result = summarizer.runInference(longText).await()

Combine this with a server-side model, and you get the same hybrid pattern as COSMO. The only difference is "who builds the Skills library."

What This Signals

What the COSMO leak shows:

1. Hybrid is becoming the standard — the single-model era is over. Local + server routing is the next baseline.
2. Skills are first-class — not chat interfaces, but bundles of triggerable capabilities are becoming the UX core.
3. Privacy trade-offs exposed — giving users a choice is the start of trust.
4. AccessibilityService redefined — screen perception is now the foundation of every mobile agent.

We don't yet know how COSMO will be officially announced at Google I/O 2026. But this architectural pattern is already standardizing.

Apple did it. Google is following. And soon, developers can build the same pattern.

Coming Next — Build It Yourself

In an upcoming post (premium tutorial series), I'll walk through building a 3-mode architecture identical to COSMO using Gemini Nano API + cloud LLM:

• AICore setup and first Nano call
• Hybrid router design (network-state-aware)
• Skills system implementation (Function Calling)
• Screen context capture (AccessibilityService)
• Privacy mode toggle UI

Before Google's official I/O announcement, getting hands-on with the underlying tech is the best preparation. Subscribe to the newsletter to be notified when it drops.

Conclusion

COSMO was an accident, but not a coincidence. It clearly shows where Google is headed:

• Agents (Skills) — chatbots → AI that acts
• Hybrid (Nano + PI) — not all-or-nothing, but routing
• Transparency (3 modes) — choice instead of automation

What gets announced at Google I/O in mid-May is worth watching. But before that, the best learning comes from building the same idea with your own hands.