Logo AgentsDroid
View GitHub Repository →
ENGINEERING RELEASE Android OS Multi-Agent Architecture

The Sovereign Multi-Agent
Runtime for Android

Transform Android into an autonomous execution platform. A central Master Agent decomposes complex natural language goals into Directed Acyclic Graphs (DAG), routing subtasks across 10 specialized micro-agents with zero mock data.

OFFICIAL SMART CONTRACT IDENTIFIER
CA SOON AVAILABLE
Verification will take place solely on this repository. Ignore unverified contract claims.
PENDING LAUNCH
10
Micro-Agents
0%
Mock Data
5 Tiers
Context Memory
< 45ms
DAG Resolution

Built for Sovereign Mobile Execution

Engineered specifically to solve real-world agent orchestration constraints on Android devices without rooting or privileged OS modifications.

Autonomous DAG Scheduler

Converts unstructured voice or text prompts into Directed Acyclic Graphs. Independent nodes run concurrently; dependent nodes wait on upstream state payloads.

MODULE DAGScheduler.kt

Native Accessibility Driver

Interacts with installed Android apps via standard AccessibilityNodeInfo hierarchies. Simulates clicks, scrolls, and text input without requiring root permissions.

SERVICE AccessibilityService

Keystore AES-256 GCM

Hardware-backed Android Keystore cryptography protects LLM credentials. Keys remain in volatile RAM only during live API invocation sessions.

SECURITY AES-256-GCM

5-Tier Memory Persistence

Organizes agent context into Sensory, Working, Episodic, Semantic, and Procedural memory tiers backed by 16 SQLite Room database entities.

STORAGE 16 Room Entities

Strict Real AI Mode

Zero fake completions or simulated mock payloads. If an endpoint is unreachable, explicit typed failure diagnostics are generated for upstream recovery.

INFERENCE 0% Synthetic

Jetpack Compose Obsidian UI

Declarative reactive user interface built strictly with modern Android Jetpack Compose, delivering fluid 120 FPS DAG graph visualization and system telemetry.

UI STACK Compose 1.6

Interactive Fleet Inspector

Explore the 10 micro-agents comprising the AgentsDroid operating runtime. Select an agent to inspect its responsibilities, permissions, and implementation code.

Master Agent

DAG Task Decomposition & Dynamic Dependency Resolver

ACTIVE SPEC
DISPATCH LATENCY
< 45ms (Resolution)
PRIMARY MODEL
GPT-4o / Claude 3.5 Sonnet
ANDROID PRIVILEGES
System Orchestration, IPC Binder
Kotlin Implementation Extract app/domain/agent/
class MasterAgent @Inject constructor(
    private val dagScheduler: DAGScheduler,
    private val memoryManager: AgentMemoryManager
) : BaseAgent() {
    suspend fun planGoal(goalPrompt: String): TaskGraph {
        val decomposedNodes = llmProvider.decomposeIntoDAG(goalPrompt)
        return dagScheduler.buildTopologicalExecutionTree(decomposedNodes)
    }
}

DAG Topological Execution Engine

Simulate how unstructured prompts are analyzed, decomposed into DAG execution nodes, and resolved sequentially across specialized agents.

READY Decomposition MasterAgent
READY Context Ingestion Research / Prod
READY Synthesis ContentAgent
READY Zero-Trust Check SecurityEngine
READY OS Dispatch Phone / Comm
Simulation runtime idle. Choose workflow and click "Execute DAG".

5-Layer Android Architecture

Engineered to conform strictly with Android OS lifecycle bounds, background execution limits, and hardware security enclaves.

LAYER 01

Jetpack Compose Obsidian UI

Declarative reactive screens rendering live telemetry, node statuses, and emergency killswitch state at 120 FPS.

Presentation
LAYER 02

Master Agent Orchestrator & DAG Scheduler

Topological graph builder, cycle detector, and concurrent task dispatcher running over Kotlin Coroutines channels.

Orchestration
LAYER 03

Zero-Trust Security & Policy Interceptor

Hardware Keystore AES-256 GCM credentials, prompt injection sanitization filters, and human approval queues.

Governance
LAYER 04

5-Tier Memory & Room Persistence

16 SQLite Room tables managing sensory, working, episodic, semantic, and procedural agent memory tiers.

Data Layer
LAYER 05

Android OS Accessibility & Intent Binder

OS-level bridge for simulating physical clicks, gestures, text input, and telephony broadcasts without root access.

OS Binder

Model Compatibility Matrix

Deploy local Wi-Fi Ollama models for 100% sovereign air-gapped execution or configure high-throughput cloud endpoints.

Provider / Engine Target Deployment Tested Models Integration Tier
Ollama Local Wi-Fi / On-Device llama3:8b, mistral:7b, phi3:mini Local Sovereign
LM Studio Local Wi-Fi Server Any OpenAI-compatible GGUF endpoint Local Compatible
Anthropic Cloud Endpoint claude-3-5-sonnet-20241022 Cloud Verified
OpenAI Cloud Endpoint gpt-4o, gpt-4o-mini Cloud Verified
Google Gemini Cloud Endpoint gemini-1.5-pro, gemini-1.5-flash Cloud Verified
Groq High-Speed Cloud llama-3.1-70b-versatile Cloud Verified
OFFICIAL SMART CONTRACT NOTICE

Token Contract Address (CA) Soon Available

The official, verified AgentsDroid contract address will only be disclosed directly via this official repository and website. Do not engage with unverified contracts, third-party direct messages, or unofficial liquidity pools claiming early access.

CA SOON AVAILABLE
Verified Repository: github.com/agentsdroid/agentsdroid

Frequently Asked Questions

No. AgentsDroid performs all device-level automation via Android's official AccessibilityService API and standard Android Intents. Root privileges are never required or requested.
Yes. By pointing AgentsDroid to an Ollama or LM Studio instance hosted on your local Wi-Fi, all reasoning, task planning, and memory operations remain strictly within your local network.
All credentials are encrypted using hardware-backed Android Keystore AES-256 GCM keys. They are decrypted strictly in volatile RAM when an outbound API request is dispatched.
The verified contract address will be placed at the very top of the repository README.md and updated on this website immediately upon launch.