Agent OS — Persistent Agent Operating System

---

name: agent-os

description: Persistent agent operating system for OpenClaw. Agents remember across sessions, learn from experience, coordinate on complex projects without duplicate work.

---

# Agent OS — Persistent Agent Operating System

Agents that remember. Learn. Coordinate.

What It Does

Agent OS enables multi-agent project execution with persistent memory:

Quick Start

Installation

clawhub install nova/agent-os

Basic Usage

const { AgentOS } = require('agent-os');

const os = new AgentOS('my-project');

// Register agents with capabilities
os.registerAgent('research', '🔍 Research', ['research', 'planning']);
os.registerAgent('design', '🎨 Design', ['design', 'planning']);
os.registerAgent('dev', '💻 Development', ['development']);

os.initialize();

// Run a project
const result = await os.runProject('Build a feature', [
  'planning',
  'design',
  'development',
]);

console.log(result.progress); // 100

Core Concepts

Agent

Persistent worker with:

TaskRouter

Decomposes goals into executable tasks:

Executor

Runs tasks sequentially:

AgentOS

Orchestrates everything:

Architecture

AgentOS (top-level orchestration)
├── Agent (persistent worker)
│   ├── Memory (lessons, capabilities, history)
│   └── State (current task, progress)
├── TaskRouter (goal decomposition)
│   ├── Templates (planning, design, development, etc.)
│   └── Matcher (task → agent assignment)
└── Executor (task execution)
    ├── Sequential runner
    ├── Progress tracking
    └── State persistence

State Persistence

All state is saved to the `data/` directory:

**This means:**

✅ Projects survive restarts

✅ Agents remember past work

✅ Resume mid-project seamlessly

File Structure

agent-os/
├── core/
│   ├── agent.js          # Agent class
│   ├── task-router.js    # Task decomposition
│   ├── executor.js       # Execution scheduler
│   └── index.js          # AgentOS class
├── ui/
│   ├── dashboard.html    # Live progress UI
│   ├── dashboard.js      # Dashboard logic
│   └── style.css         # Styling
├── examples/
│   └── research-project.js  # Full working example
├── data/                 # Auto-created (persistent state)
└── package.json

API Reference

AgentOS

new AgentOS(projectId?)
registerAgent(id, name, capabilities)
initialize()
runProject(goal, taskTypes)
getStatus()
getAgentStatus(agentId)
toJSON()

Agent

startTask(task)
updateProgress(percentage, message)
completeTask(output)
setBlocker(message)
recordError(error)
learnLesson(category, lesson)
reset()
getStatus()

TaskRouter

decompose(goal, taskTypes)
matchAgent(taskType)
getTasksForAgent(agentId, tasks)
canExecuteTask(task, allTasks)
getNextTask(tasks)
completeTask(taskId, tasks, output)
getProjectStatus(tasks)

Executor

initializeProject(goal, taskTypes)
execute()
executeTask(task)
getStatus()

Example: Research + Design + Development

See `examples/research-project.js` for the canonical example:

npm start

This demonstrates:

Expected output:

✅ Registered 3 agents
📋 Task Plan: 12 tasks
🚀 Starting execution...
✅ [Task 1] Complete
✅ [Task 2] Complete
...
📊 PROJECT COMPLETE - 100% progress

What's Coming (v0.2+)

Philosophy

**Agents should remember what they learn.**

Most agent frameworks are stateless. Agent OS keeps persistent memory so agents:

1. **Remember** — No redundant context resets

2. **Learn** — Capability scores improve over time

3. **Coordinate** — Shared state prevents duplication

4. **Cost less** — Less context = cheaper API calls

License

MIT

---

**Built with ❤️ by Nova for OpenClaw**

See README.md and ARCHITECTURE.md for complete documentation.