Google’s COSMO Leak Exposes Shift to On-Device AI Agents

Google slipped up in early May. The company published an experimental Android app called COSMO to the Play Store under its main account. Then it yanked the listing within hours. What leaked offers a rare look at how Google plans to move artificial intelligence from distant servers onto phones themselves.

The app weighed in at 1.13 GB. That size comes from bundling a full copy of Gemini Nano, Google’s compact model designed to run locally. Unlike the standard Gemini app, which ships most queries to the cloud, COSMO can operate offline. It can also tap stronger remote models when connected. The brief appearance, first spotted by 9to5Google, revealed far more than a simple chatbot.

COSMO requests broad system permissions. It asks for access to screen content through Android’s AccessibilityService. The goal appears straightforward. The assistant watches what users do across apps. It listens to conversations. It suggests actions before anyone types a prompt. A text thread about dinner plans might trigger a calendar entry offer. Mention a document and it volunteers to draft one.

Fourteen distinct skills point to an agent that acts, not just answers.

Those skills read like a list of everyday frustrations solved. List Tracker suggests shopping lists when plans form. Document Writer generates drafts or summaries on demand. Calendar Event Suggester detects agreed times in messages and books them. Browser Agent, powered by an internal tool called Mariner, completes web tasks on the user’s behalf. Add Timer sets clock reminders for time-bound activities. Deep Research compiles reports from multiple sources when simple answers fall short.

Quick Photo Lookup pulls specific images from the gallery without breaking conversation flow. Google it offers to search the web. Jargon Definitions explains acronyms or technical terms on the fly. Provide Insight adds background during reading or discussion. People Understanding and Event Understanding supply context about contacts or occasions. Recall surfaces forgotten details. Conversation Summary condenses recent talks when users switch tasks.

Ben Khalesi at Android Police captured the appeal. “I’m officially burnt out on chatbots,” he wrote. “Typing prompts into a text box feels like learning a programming language to do the work my phone should already understand.” COSMO, in his view, represents the first mobile AI concept in some time that holds real promise.

The settings menu offered three fulfillment modes. Hybrid, the presumed default, uses a server-side model called PI when online and falls back to Nano offline. PI Only sticks to the remote system for maximum capability. Nano Only confines everything to the device. “PI” remains unexplained in the leaked materials, though observers link it to personal intelligence or a more advanced cloud model.

Privacy questions follow immediately. Accessibility access can expose sensitive on-screen information. In a purely cloud-based system that would raise alarms. Local processing with Gemini Nano keeps much of the analysis on the device. Data never leaves the phone for certain tasks. Battery life becomes the tradeoff. Running even a small model continuously can drain power if optimization lags. Khalesi noted the concern but found the local approach made the permission feel less invasive than it might otherwise.

Google built COSMO under its research umbrella. The package name com.google.research.air.cosmo signals origins in the company’s advanced projects group. The description on the short-lived Play Store page promised an assistant that “brings the power of artificial intelligence directly onto your device” and “works behind the scenes to simplify your life.” Screenshots showed a bare-bones chat interface. The experimental tag was clear from the start.

Analysts see COSMO as a test bed rather than a consumer launch. Features that prove effective could migrate into the main Gemini app or deeper Android system integration. The timing, weeks before Google I/O, fueled speculation that elements were destined for the keynote stage. The accidental release instead gave outsiders an unfiltered preview.

This focus on local models aligns with broader industry moves. Apple has pushed on-device intelligence in recent iOS releases. Qualcomm and Arm continue to optimize silicon for AI inference. Google itself has invested heavily in making Gemini Nano efficient enough for phones. The 1.13 GB footprint shows the current cost of that capability. Future iterations will likely shrink.

Connection to Project Astra adds another layer. Google’s DeepMind research project has explored universal assistants that maintain context across devices and understand the physical world through cameras. While COSMO operates within the phone’s screen and apps rather than augmented reality glasses, the shared ambition stands out. Both efforts aim to create systems that observe, remember, and act with minimal user direction.

Recent coverage reinforces the experimental status. Android Headlines described COSMO as offering a glimpse of deeper, more private AI integration in Android. The publication highlighted proactive agent behavior over reactive prompting. Droid Life reported the app could crawl into nearly all aspects of the device, raising both excitement and caution about always-listening capabilities.

Users who installed the app before removal described a basic interface with voice match setup and screen access toggles. Performance details remain limited. Early testers noted the local model handled simple queries quickly but leaned on cloud resources for complex research or browser automation. Hybrid mode appeared to balance speed, privacy, and capability in early hands-on reports shared on forums.

The leak arrives at a moment when consumer fatigue with generative AI chatbots has grown. Many tools demand precise instructions and still hallucinate or miss context. An assistant that reads the screen, understands conversation flow, and volunteers relevant actions could cut through that friction. Yet success depends on execution. Too many false suggestions will annoy users. Insufficient privacy safeguards will drive them away.

Google has not commented publicly on COSMO since pulling the listing. Company spokespeople declined to elaborate when contacted by multiple outlets. The silence leaves room for interpretation. Perhaps the app represented an early prototype. Or maybe it contained capabilities the company preferred to refine before wider exposure.

One thing seems certain. The direction points toward agents that live on the device. They will combine local speed and privacy with selective cloud power. They will watch context across messaging, calendars, photos, and browsers. They will summarize, suggest, research, and automate. The question is how quickly Google can polish these abilities and how users will respond to an assistant that sees so much.

Khalesi ended his piece on a measured note. Nobody should treat the leaked version as a finished product. He expects Google to fold successful elements into Gemini or Android itself over time. “With Google I/O getting close,” he added, “I can’t help but wonder if COSMO was headed for the keynote. I’m definitely curious to see what Google says next.”

The rest of the industry watches too. On-device AI has moved from research curiosity to practical prototype. COSMO, however briefly, showed what that future might feel like on a phone in your pocket. Less typing. More understanding. And a careful balance between capability and control.