Skip to main content
Hands are autonomous background workers that run on schedules to perform research, monitoring, data collection, and workflow automation without human intervention. Unlike interactive agents that respond to messages, hands proactively execute multi-phase workflows on cron schedules and store their results in memory for later retrieval.

​
Quick Start

# List available hands
agentos hand list

# Enable a hand
agentos hand enable researcher

# Configure hand settings
agentos hand configure researcher --set research_depth=thorough

# View hand metrics
agentos hand status researcher

# Trigger manual execution
agentos hand run researcher --input '{"topic": "quantum computing trends 2026"}'

​
Available Hands

AgentOS includes 7 specialized hands for common automation needs:

Researcher

Deep research automation that finds, cross-references, and synthesizes information.
  • Schedule: Every 6 hours (0 */6 * * *)
  • Phases: 7-phase research workflow
  • Output: Structured reports with citations
agentos hand enable researcher

Browser

Web automation agent for form filling, data extraction, and multi-step web workflows.
  • Schedule: Every 2 hours (0 */2 * * *)
  • Phases: Navigation, interaction, extraction
  • Output: Extracted data and task completion logs
agentos hand enable browser

Collector

Data gathering worker that pulls data from APIs, databases, and web sources on schedule.
  • Schedule: Configurable (default: hourly)
  • Phases: Source polling, validation, storage
  • Output: Normalized datasets in memory
agentos hand enable collector

Predictor

Forecasting agent that analyzes trends and predicts future values.
  • Schedule: Daily at midnight (0 0 * * *)
  • Phases: Data collection, model training, prediction
  • Output: Time-series forecasts with confidence intervals
agentos hand enable predictor

Twitter

Social media monitor for mentions, trends, and engagement tracking.
  • Schedule: Every 15 minutes (*/15 * * * *)
  • Phases: Scanning, sentiment analysis, response drafting
  • Output: Queued responses and analytics
agentos hand enable twitter

Clip

Content extraction worker that saves articles, videos, and podcasts for later processing.
  • Schedule: Continuous monitoring
  • Phases: Discovery, extraction, summarization
  • Output: Archived content with metadata
agentos hand enable clip

Lead

Sales lead monitor that tracks prospects and qualifies leads automatically.
  • Schedule: Every 4 hours (0 */4 * * *)
  • Phases: Source scanning, qualification, CRM update
  • Output: Qualified lead list with scores
agentos hand enable lead

​
Deep Dive: Researcher Hand

The researcher hand executes a 7-phase autonomous research workflow:
1

Topic Decomposition

Breaks the research query into 3-5 sub-questions covering the full scope. Identifies key terms, related concepts, and potential angles.
2

Source Discovery

Searches for high-quality sources across academic papers, news, documentation, and expert blogs. Targets the configured source count (default: 10).
3

Information Extraction

Reads each source and extracts facts, statistics, quotes, and arguments. Notes contradictions and flags unverified claims.
4

Cross-Reference Validation

Compares findings across sources. Identifies consensus views, disputed points, and gaps. Assigns confidence levels.
5

Synthesis

Organizes validated findings into a coherent narrative with introduction, key findings, analysis, and conclusions. Includes citations.
6

Memory Storage

Stores the completed report in memory with tags, source URLs, and metadata. Links to related previous research.
7

Follow-Up Identification

Lists unanswered questions and suggests follow-up topics. If auto_follow_up is enabled, queues the top follow-up for next run.

​
Configuration

[hand]
id = "researcher"
name = "Deep Researcher"
description = "Autonomously researches topics and stores structured reports"
enabled = true
schedule = "0 */6 * * *"

[hand.tools]
allowed = [
  "tool::web_search",
  "tool::web_fetch",
  "tool::file_write",
  "memory::store",
  "memory::query"
]

[[hand.settings]]
key = "research_depth"
type = "select"
default = "thorough"
options = ["quick", "standard", "thorough", "exhaustive"]

[[hand.settings]]
key = "source_count"
type = "number"
default = "10"
options = ["5", "10", "20", "50"]

[hand.agent]
max_iterations = 80
temperature = 0.3

​
Usage Example

# Enable researcher hand with custom settings
agentos hand enable researcher
agentos hand configure researcher \
  --set research_depth=exhaustive \
  --set source_count=20 \
  --set auto_follow_up=true

# Queue a research topic
agentos hand run researcher --input '{
  "topic": "Rust async runtime performance comparison 2026",
  "tags": ["rust", "performance", "async"]
}'

# View results
agentos memory list researcher
agentos memory get researcher latest_report

​
Deep Dive: Browser Hand

The browser hand performs headless browser automation for complex multi-step web tasks:
1

Task Parsing

Analyzes the browsing task to identify target website, goal actions, and expected outcomes. Plans navigation path and identifies barriers.
2

Navigation

Opens target URL, waits for full page load, takes verification screenshot. Handles redirects, cookie banners, and popups automatically.
3

Page Analysis

Extracts page structure including interactive elements (buttons, forms, links, dropdowns). Maps element selectors for reliable interaction.
4

Interaction Execution

Performs actions (clicking, typing, selecting) in correct sequence. Waits for page transitions and verifies each action before proceeding.
5

Data Extraction

Parses relevant page content into structured format. Handles pagination by detecting and following next-page links.
6

Error Recovery

On failure, takes screenshot, analyzes page state, attempts alternative approach. Retries up to 3 times with increasing wait intervals.
7

Result Storage

Stores extracted data and completion status in memory. Saves screenshots of key states for audit trail. Generates task summary.

​
Configuration

[hand]
id = "browser"
name = "Web Browser Agent"
enabled = true
schedule = "0 */2 * * *"

[[hand.settings]]
key = "headless_mode"
type = "boolean"
default = "true"

[[hand.settings]]
key = "page_timeout_ms"
type = "number"
default = "30000"
options = ["10000", "30000", "60000"]

[[hand.settings]]
key = "max_pages_per_task"
type = "number"
default = "20"

​
Usage Example

# Enable browser automation
agentos hand enable browser

# Configure for visible browser (debugging)
agentos hand configure browser --set headless_mode=false

# Queue a scraping task
agentos hand run browser --input '{
  "task": "Extract product prices from ecommerce site",
  "url": "https://example.com/products",
  "selectors": {
    "product": ".product-card",
    "price": ".price",
    "name": ".product-title"
  },
  "pagination": true
}'

​
Deep Dive: Twitter Hand

The Twitter hand monitors social media for brand mentions, trends, and engagement opportunities:
1

Mention Scanning

Searches Twitter/X for mentions of tracked keywords, handles, and hashtags. Collects tweet text, author info, engagement metrics, and thread context.
2

Sentiment Analysis

Classifies each mention as positive, negative, neutral, or question. Applies sentiment filter. Scores urgency based on author reach and negative intensity.
3

Thread Reconstruction

For conversation tweets, reconstructs full thread to understand context. Identifies original poster, key points, and directed questions.
4

Response Drafting

Drafts replies matching configured tone. Keeps responses concise. For negative mentions, acknowledges concerns first. For questions, provides direct answers with links.
5

Content Opportunity Detection

Identifies trending topics relevant to tracked keywords. Drafts original tweet suggestions with hashtags and optimal posting times.
6

Queue Management

Stores drafted responses with status (pending/approved/posted). In draft mode, holds for review. In auto_reply mode, posts below hourly cap. In notify_only, just stores summaries.
7

Analytics Update

Updates engagement metrics, response rates, sentiment trends, and follower interaction patterns. Generates summary of key conversations and trends.

​
Configuration

[[hand.settings]]
key = "track_keywords"
type = "text"
default = ""

[[hand.settings]]
key = "response_mode"
type = "select"
default = "draft"
options = ["draft", "auto_reply", "notify_only"]

[[hand.settings]]
key = "max_responses_per_hour"
type = "number"
default = "10"

[[hand.settings]]
key = "tone"
type = "select"
default = "professional"
options = ["professional", "casual", "witty", "technical"]

​
Creating Custom Hands

Create a new hand by defining a TOML configuration: 
[hand]
id = "stock-monitor"
name = "Stock Market Monitor"
description = "Tracks stock prices and alerts on significant moves"
enabled = true
schedule = "*/30 * * * *"  # Every 30 minutes

[hand.tools]
allowed = [
  "tool::web_fetch",
  "tool::http_request",
  "memory::store",
  "memory::query"
]

[[hand.settings]]
key = "watch_list"
type = "text"
default = "AAPL,GOOGL,MSFT"

[[hand.settings]]
key = "alert_threshold_pct"
type = "number"
default = "5"
options = ["1", "3", "5", "10"]

[hand.agent]
max_iterations = 40
temperature = 0.1
system_prompt = """
Phase 1 - Price Collection: Fetch current prices for all symbols in watch_list.
Phase 2 - Historical Comparison: Compare to previous run stored in memory.
Phase 3 - Alert Detection: Flag any moves exceeding alert_threshold_pct.
Phase 4 - Analysis: For flagged stocks, analyze news and sentiment.
Phase 5 - Notification: Send alerts for significant moves with context.
Phase 6 - Memory Update: Store latest prices and analysis for next run.
"""

[[hand.dashboard.metrics]]
label = "Stocks Tracked"
key = "stocks_count"

[[hand.dashboard.metrics]]
label = "Alerts Triggered"
key = "alerts_count"

​
Hand Management

agentos hand list
agentos hand list --enabled

agentos hand enable researcher
agentos hand disable browser

agentos hand status researcher
agentos hand metrics researcher --last 24h

# Run immediately with input
agentos hand run researcher --input '{"topic": "AI trends"}'

# Run with custom settings override
agentos hand run researcher \
  --input '{"topic": "Rust"}' \
  --override research_depth=quick

agentos hand logs researcher
agentos hand logs researcher --last 10
agentos hand logs researcher --since "2026-03-01"

​
Dashboard Metrics

Hands expose custom metrics visible in the TUI dashboard: 
[[hand.dashboard.metrics]]
label = "Reports Generated"
key = "reports_count"

[[hand.dashboard.metrics]]
label = "Sources Analyzed"
key = "sources_analyzed"

[[hand.dashboard.metrics]]
label = "Avg Confidence Score"
key = "avg_confidence"
Access metrics programmatically: 
agentos hand metrics researcher --format json

​
Integration with Workflows

Hands can be invoked as workflow steps: 
steps:
  - name: Background Research
    functionId: "hand::run"
    mode: sequential
    timeoutMs: 300000
    errorMode: retry
    maxRetries: 2
    outputVar: research_output
    config:
      handId: researcher
      input:
        topic: "{{workflow_input.topic}}"
        tags: ["automated"]

​
Security & Approval Gates

Hands respect the same security model as agents:
  • RBAC: Tool access controlled via [hand.tools].allowed
  • Approval Tiers: Destructive actions require async/sync approval
  • Sandboxing: Code execution runs in Docker/WASM sandbox
  • Rate Limiting: Configurable per-hand rate limits
  • Audit Trail: All actions logged to Merkle chain
[hand.security]
approval_tier = "async"  # auto | async | sync
max_requests_per_hour = 100
sandbox = "docker"  # docker | wasm | none

​
Best Practices

Idempotency

Design hands to be idempotent — running multiple times should not cause duplicate work or data corruption.

Error Recovery

Store progress in memory so hands can resume from last checkpoint after failures.

Resource Limits

Set max_iterations and timeouts to prevent runaway execution and cost overruns.

Observability

Emit custom metrics for monitoring hand health and effectiveness over time.

​
Next Steps

Workflows

Combine hands with agents in multi-step workflows

Memory

Store and retrieve hand results using the memory system

Scheduling

Configure cron schedules and pulse-based triggers

Approval System

Control hand actions with approval gates

Build docs developers (and LLMs) love

Get started for freeTalk to us