OSH-2026 Agent Runtime: A Structured Execution System for Android On-Device Agents

This article introduces the OSH-2026 Agent Runtime, a course project from the University of Science and Technology of China's OSH-2026 program. It is an agent execution system deployable locally on Android, with core components including a DAG-driven Action Fabric execution framework and on-device LLM inference capabilities. It enables a complete pipeline from model planning to system tool execution, aiming to build a locally deployable, structured-execution, observable, and extensible on-device agent technology stack.

Original Authors/Maintainers: OSH-2026 Course Organization Source: GitHub (Link: https://github.com/OSH-2026/agent_runtime) Release Date: June 12, 2026 This project is a major assignment for the OSH-2026 course, aiming to build a locally deployable, structured-execution, observable, and extensible Agent Runtime for Android on-device agents. It centers on two subsystems: Action Fabric (structured execution system) and the on-device LLM inference framework, forming a complete technology stack.

Action Fabric: Explicit DAG Execution System

Traditional agent execution processes are implicit in dialogue contexts, making them difficult to track and debug. Action Fabric innovatively represents the execution process explicitly as a Directed Acyclic Graph (DAG). It validates dependencies before execution, computes the set of ready nodes at runtime, and uniformly controls state, error recovery, and side effects.

Rust Scheduling Kernel Capabilities

• DAG Construction and Validation: Load ActionFlow YAML, automatically establish dependencies, detect duplicate nodes, cyclic dependencies, etc.;
• Scheduling Strategy: Compute the set of ready nodes, support batch asynchronous execution of independent nodes, and apply serial constraints to non-idempotent Actions;
• Fault Tolerance and Recovery: Support risk levels, confirmation gating, timeout retries, implement bounded recovery mechanisms, and record state transitions and audit logs.

The Kotlin Runtime has implemented and registered 59 Android Actions, covering device status, app management, network, file, media, and other categories. Examples include:

Category	Example Action
Device & System	device_info, power_status
Network & Connection	network_status, wifi_toggle
App & Intent	launch_app, intent_show_map
Data & File	read_file, clipboard_copy
Personal Info	search_contacts, read_sms
Media & Sensors	take_photo, screenshot
System Interaction	set_alarm, media_play_pause
The Android Runtime runs a gRPC Server as a foreground service, with supporting functions such as permission application and Intent Host.

Cross-Language Execution Chain

Rust-Kotlin communication is implemented via Protocol Buffers and gRPC:

1. Rust RemoteAction forwards tool nodes to the Kotlin Runtime;
2. On the Kotlin side, input decoding, execution, and result encoding are completed via ActionExecutor + JsonCodec + ActionRegistry;
3. The network layer uses Rustls, and cross-compilation verification for Android aarch64 has been completed.

Tauri Demo Application

A Tauri 2 Android application is provided, supporting: editing ActionFlow YAML, automatic DAG scheduling parsing, calling local/LAN Kotlin Runtime, displaying execution status and results, etc., to verify the complete scenario pipeline.

OSH-2026 Agent Runtime demonstrates an engineering path for on-device agents:

1. Structured Execution: Convert implicit processes to explicit DAGs, improving observability and controllability;
2. On-Device Deployment: Local operation protects user privacy;
3. Extensible Architecture: Registry mechanism supports dynamic addition of new Actions;
4. Cross-Language Collaboration: Rust's high-performance scheduling complements Kotlin's Android ecosystem. This project provides a reference architecture paradigm for on-device agent engineering implementation, with practical value in execution reliability, error recovery, and system capability invocation.