Tech • IA • Crypto
Create agent skills for on-device generative AI (I/O Connect ‘26)
8/10TL;DR
Google showcased how developers can build “agent skills” to extend on-device AI models using its AI Edge stack, enabling faster, private, and offline-capable applications.
KEY POINTS
Push toward on-device AI
Google is advancing AI Edge, a cross-platform toolkit designed to run machine learning models directly on devices including Android, iOS, web browsers, desktops, and IoT hardware. Running models locally reduces latency by eliminating network delays, improves privacy by keeping data on-device, and enables full offline functionality. It also lowers infrastructure costs by shifting computation away from cloud data centers.
End-to-end deployment pipeline
The AI Edge stack provides a full workflow from raw model to deployment. Tools such as LiteRTorch convert models into compatible formats, while Edge Quantizer reduces model size to fit device constraints without major performance loss. The runtime layer, LiteRT, supports execution across CPUs, GPUs, and NPUs, ensuring efficient performance on diverse hardware.
Dedicated APIs for generative AI
To address the complexity of running large language models locally, Google introduced LiteTLM APIs, which handle tasks such as KV caching, context window management, and text generation loops. These abstractions allow developers to focus on user experience instead of low-level optimization, making on-device generative AI more accessible.
Cross-platform and developer-friendly design
The framework supports development across multiple environments with Kotlin and Swift bindings, Flutter plugins, and lower-level C++ APIs for advanced use cases. This flexibility enables developers to build once and deploy broadly, from smartphones to edge devices like Raspberry Pi.
Agent skills extend model capabilities
A central concept is the use of agent skills, a structured way to give models new abilities. Skills reduce repetitive prompting, enable dynamic capability discovery, and inject domain-specific knowledge. They are particularly useful for smaller on-device models that mayحتاج more guidance to perform complex tasks accurately.
Structured skill architecture
Each skill is packaged as a directory containing metadata, instructions, executable code, and assets. A markup file defines the skill’s identity, description, and operational rules, while additional components handle logic and user interface rendering. This modular design allows skills to function like plug-ins that models can invoke when needed.
Real-world demo: task planner and timer
A demonstration showcased a “task planner and timer” skill that compensates for LLM limitations in understanding time. When prompted with multiple activities, the model automatically invoked the skill, generated structured function calls, and rendered a UI with active timers. The example highlighted how deterministic code and UI components can augment generative outputs.
Testing and optimization tools
Developers can validate performance using Edge Portal, which benchmarks models across real devices, and Model Explorer, which helps identify architectural bottlenecks. These tools aim to ensure reliable deployment in real-world conditions before release.
Open-source ecosystem and experimentation
Google emphasized that its AI Edge Gallery app and related tools are open source, allowing developers to inspect implementations and reuse components. Integration with coding agents can further accelerate development by generating boilerplate for new skills.
CONCLUSION
Google’s AI Edge framework and agent skills approach signal a broader shift toward powerful, customizable AI running directly on devices, combining efficiency, privacy, and extensibility for next-generation applications.