Skip to content

AGENTS.md

Latest commit

 

History

History
223 lines (175 loc) · 5.8 KB

AGENTS.md

File metadata and controls

223 lines (175 loc) · 5.8 KB

NeuroOS Agents Documentation

Overview

NeuroOS features a powerful multi-agent AI system inspired by Crew AI, designed to handle complex tasks through collaborative agent workflows. Each agent has specific roles, expertise, and access to system tools.

Agent System Architecture

Agent Roles

Agent Role Description
Coordinator Orchestrator Manages workflow, delegates tasks, synthesizes results
Planner Strategist Analyzes tasks, creates execution plans, maps dependencies
Researcher Information Gatherer Browses web, scrapes pages, retrieves data
Executor Action Performer Creates files, runs commands, executes tasks
Analyst Reviewer Analyzes data, reviews outputs, provides insights

Available Tools by Agent

Coordinator Agent

  • open_app - Open system applications
  • close_app - Close running applications
  • list_running_apps - List open windows
  • send_notification - Show system notifications
  • get_system_info - Get system information
  • open_file - Open files in viewer
  • add_board_widget - Add widgets to NeuroBoard
  • update_memory - Save persistent memory
  • save_to_workspace - Save files to workspace

Planner Agent

  • open_app - Open applications
  • list_running_apps - List running apps
  • get_system_info - System information
  • get_app_windows - Get window list
  • update_memory - Update memory
  • list_files - List workspace files
  • read_file - Read file contents

Researcher Agent

Browser Tools:

  • browser_navigate - Navigate to URL
  • browser_scrape - Extract page content
  • web_fetch - CORS-free fetch
  • search_web - Google search
  • web_research - Multi-URL research
  • browser_click - Click elements
  • browser_type - Type in inputs
  • browser_key - Press keyboard keys
  • browser_submit - Submit forms
  • browser_scroll - Scroll pages
  • browser_wait - Wait for elements
  • browser_evaluate - Execute JavaScript
  • browser_get_info - Get page metadata
  • browser_get_links - Extract links
  • browser_get_html - Get raw HTML
  • browser_save - Save content to workspace

File Tools:

  • read_file - Read files
  • list_files - List directory contents

Executor Agent

OS Tools:

  • open_app, close_app, list_running_apps
  • send_notification, open_file
  • add_board_widget, update_memory, save_to_workspace

File Tools:

  • save_file - Create/write files
  • read_file - Read files
  • append_file - Append to files
  • update_file - Update files (find/replace)
  • list_files - List directory
  • create_folder - Create directories
  • delete_file - Delete files

Shell Tools:

  • run_shell - Execute shell commands

Generate Tools:

  • generate_report - Generate markdown reports
  • generate_image - Generate AI images

Analyst Agent

  • get_app_windows, get_system_info, list_running_apps
  • update_memory
  • list_files, read_file, save_file, append_file
  • browser_scrape, web_fetch, browser_get_info, browser_get_links
  • generate_report

Using the Crew AI System

Basic Usage

import { createCrew } from './lib/ai/crew';

const crew = createCrew({
    tasks: [
        { 
            id: 'research', 
            description: 'Research the latest AI trends', 
            status: 'pending', 
            dependencies: [] 
        },
        { 
            id: 'write_report', 
            description: 'Write a research report', 
            status: 'pending', 
            dependencies: ['research'] 
        }
    ],
    verbose: true,
    planningEnabled: true
});

const result = await crew.execute(context);
console.log(result.finalOutput);

Task Configuration

interface Task {
    id: string;
    description: string;
    status: 'pending' | 'in_progress' | 'completed' | 'failed';
    assignedAgent?: string;
    result?: string;
    error?: string;
    dependencies: string[];
}

Execution Flow

  1. Planning Phase - Planner analyzes tasks and assigns agents
  2. Execution Phase - Agents execute their assigned tasks
  3. Dependency Resolution - Tasks wait for dependencies to complete
  4. Synthesis - Coordinator aggregates results

Browser Automation

The Researcher agent has full access to Neuro Browser automation:

Navigation

{ "tool": "browser_navigate", "args": { "url": "https://example.com" } }

Scraping

{ "tool": "browser_scrape", "args": { "maxLength": 6000 } }

Search

{ "tool": "search_web", "args": { "query": "latest AI news" } }

Interaction

{ "tool": "browser_click", "args": { "selector": "#submit-btn" } }
{ "tool": "browser_type", "args": { "text": "hello", "selector": "#input" } }

File Operations

Read File

{ "tool": "read_file", "args": { "filename": "data.json" } }

Write File

{ "tool": "save_file", "args": { "filename": "output.md", "content": "# Report" } }

List Files

{ "tool": "list_files", "args": { "path": "documents" } }

Best Practices

  1. Clear Task Descriptions - Provide detailed task descriptions for better results
  2. Set Dependencies - Define task dependencies for proper execution order
  3. Use Appropriate Agents - Match tasks to agent capabilities
  4. Enable Planning - Let the Planner optimize task assignments
  5. Monitor Execution - Use verbose mode to track progress

Error Handling

The system includes:

  • Automatic retry for transient failures
  • Dependency deadlock detection
  • Maximum iteration limits (prevents infinite loops)
  • Timeout handling for long-running operations

Performance

  • Task execution is parallelized when dependencies allow
  • Tool results are cached for reuse
  • Memory is persisted across sessions
  • Browser automation uses efficient DOM targeting