Human-in-the-Loop Integration

Learn how to integrate human oversight and collaboration into your AI agent workflows

Human-in-the-Loop Integration

RCrewAI provides comprehensive human-in-the-loop functionality that allows humans to collaborate with AI agents, providing oversight, guidance, and intervention when needed.

Overview

Human-in-the-loop capabilities include:

Task Confirmation: Human approval before starting critical tasks
Tool Approval: Human confirmation for potentially sensitive tool usage
Real-time Guidance: Human input during agent reasoning processes
Result Review: Human review and feedback on task completion
Error Recovery: Human intervention when agents encounter failures

Basic Human-in-the-Loop Setup

require 'rcrewai'

# Configure RCrewAI
RCrewAI.configure do |config|
  config.llm_provider = :openai
end

# Create an agent with human interaction enabled
agent = RCrewAI::Agent.new(
  name: "research_agent",
  role: "Research Specialist", 
  goal: "Conduct thorough research with human oversight",
  backstory: "Expert researcher who works closely with humans for quality assurance",
  tools: [RCrewAI::Tools::WebSearch.new, RCrewAI::Tools::FileWriter.new],
  human_input: true,                      # Enable human interaction
  require_approval_for_tools: true,       # Require approval for tool usage
  require_approval_for_final_answer: true # Require approval for final results
)

# Create a task with human interaction points
task = RCrewAI::Task.new(
  name: "sensitive_research", 
  description: "Research confidential market information for our new product",
  agent: agent,
  expected_output: "Market analysis report with strategic insights",
  human_input: true,                # Enable human interaction for this task
  require_confirmation: true,       # Human must confirm before starting
  allow_guidance: true,            # Allow human guidance during execution
  human_review_points: [:completion] # Review when task completes
)

# Execute - human will be prompted for input as needed
result = task.execute

Human Interaction Types

1. Task Confirmation

Tasks can require human confirmation before starting:

task = RCrewAI::Task.new(
  name: "deploy_changes",
  description: "Deploy the new model to production",
  require_confirmation: true,
  agent: deployment_agent
)

# When executed, human will see:
# 🤝 HUMAN APPROVAL REQUIRED
# ========================================
# Request: Confirm execution of task: deploy_changes
# Context: Description: Deploy the new model to production
# Expected Output: Not specified
# Assigned Agent: deployment_agent
# 
# Consequences: The task will be executed with the specified agent and may use external tools.
# 
# Do you approve this action? (yes/no)

2. Tool Approval Workflows

Agents can request human approval before using potentially sensitive tools:

agent = RCrewAI::Agent.new(
  name: "data_agent",
  role: "Data Processor",
  goal: "Process data safely",
  tools: [RCrewAI::Tools::SqlDatabase.new, RCrewAI::Tools::EmailSender.new],
  require_approval_for_tools: true
)

# When agent wants to use a tool, human will see:
# 🤝 HUMAN APPROVAL REQUIRED  
# ========================================
# Request: Agent data_agent wants to use tool 'SqlDatabase'
# Context: Parameters: {"query": "SELECT * FROM customers", "database": "production"}
# Consequences: This will execute the SqlDatabase tool with the specified parameters.
#
# Do you approve this action? (yes/no)

3. Real-time Guidance During Reasoning

Humans can provide guidance during agent reasoning loops:

agent = RCrewAI::Agent.new(
  name: "strategic_planner",
  role: "Business Strategist",
  goal: "Develop strategic plans with human input",
  human_input: true
)

task = RCrewAI::Task.new(
  name: "market_strategy",
  description: "Develop go-to-market strategy for new product",
  agent: agent,
  allow_guidance: true
)

# During execution, human will periodically see:
# 👀 HUMAN REVIEW REQUESTED
# ========================================
# Content to review:
# Task: market_strategy
# Description: Develop go-to-market strategy for new product
# 
# Current Iteration: 1
# 
# Agent Analysis:
# - Role: Business Strategist
# - Current Progress: Starting task
# - Previous Reasoning: No previous reasoning
# 
# The agent is about to continue reasoning for this task.
# ----------------------------------------
# 
# Review criteria: Task approach, Progress assessment, Strategic guidance
# 
# Please review and provide feedback (or type 'approve' to approve as-is):

4. Final Answer Review and Revision

Agents can request human review of their final answers:

agent = RCrewAI::Agent.new(
  name: "content_creator",
  role: "Content Writer", 
  goal: "Create high-quality content",
  require_approval_for_final_answer: true
)

# When task completes, human will see:
# 👀 HUMAN REVIEW REQUESTED
# ========================================
# Content to review:
# [Agent's final answer/result here]
# ----------------------------------------
# 
# Review criteria: Accuracy, Completeness, Clarity, Relevance
# 
# Please review and provide feedback (or type 'approve' to approve as-is):

# If human provides feedback, agent can:
# 1. Revise the answer based on feedback
# 2. Use the answer as-is despite feedback  
# 3. Let human provide the correct answer

5. Error Recovery with Human Intervention

When agents encounter failures, humans can guide the recovery process:

# If a task fails, human will see:
# 🎯 HUMAN CHOICE REQUIRED
# ========================================
# Question: Task 'data_analysis' failed with error: Connection timeout. How should I proceed?
# 
# Available choices:
#   1. Retry with current settings
#   2. Modify task parameters and retry
#   3. Abort task execution
# 
# Please select a choice (enter number or text):

# If human selects "2. Modify task parameters":
# 💬 HUMAN INPUT REQUESTED
# ========================================
# Prompt: Please specify task modifications (JSON format):
# 
# Help: Provide modifications as JSON, e.g. {"description": "new description", "expected_output": "new output"}
# 
# Please provide your input:

Advanced Human Interaction Patterns

Multi-Agent Collaboration with Human Oversight

# Create crew with manager coordination
crew = RCrewAI::Crew.new("human_assisted_team", process: :hierarchical)

# Manager agent coordinates human interactions
manager = RCrewAI::Agent.new(
  name: "project_manager",
  role: "Project Coordinator", 
  goal: "Coordinate team with human oversight",
  manager: true,
  human_input: true
)

# Specialist agents with different human interaction needs
researcher = RCrewAI::Agent.new(
  name: "researcher",
  role: "Research Specialist",
  goal: "Conduct research with human guidance",
  human_input: true,
  require_approval_for_tools: true  # Requires approval for web searches
)

analyst = RCrewAI::Agent.new(
  name: "analyst", 
  role: "Data Analyst",
  goal: "Analyze data with human validation",
  human_input: true,
  require_approval_for_final_answer: true  # Human validates analysis results
)

crew.add_agent(manager)
crew.add_agent(researcher) 
crew.add_agent(analyst)

# Tasks with different human interaction requirements
research_task = RCrewAI::Task.new(
  name: "market_research",
  description: "Research competitor landscape",
  human_input: true,
  require_confirmation: true,  # Confirm before starting research
  human_review_points: [:completion]
)

analysis_task = RCrewAI::Task.new(
  name: "competitive_analysis", 
  description: "Analyze research findings for strategic insights",
  context: [research_task],
  human_input: true,
  allow_guidance: true,        # Allow human strategic input
  human_review_points: [:completion]
)

crew.add_task(research_task)
crew.add_task(analysis_task)

# Execute with human collaboration
results = crew.execute

Custom Human Input Utilities

# Use HumanInput directly for custom interactions
human_input = RCrewAI::HumanInput.new(verbose: true)

# Request approval with context
approval = human_input.request_approval(
  "Deploy model to production environment",
  context: "Model version: v2.1.4, Environment: Production",
  consequences: "This will update the live model serving production traffic",
  timeout: 120
)

if approval[:approved]
  puts "Deployment approved: #{approval[:reason]}"
else
  puts "Deployment rejected: #{approval[:reason]}"
end

# Request multiple choice input
choice = human_input.request_choice(
  "Which deployment strategy should we use?",
  ["Blue-green deployment", "Rolling deployment", "Canary deployment"],
  timeout: 60
)

puts "Selected strategy: #{choice[:choice]}" if choice[:valid]

# Request text input with validation
input = human_input.request_input(
  "Enter the deployment configuration:",
  type: :json,
  validation: {
    required_keywords: ["environment", "replicas"],
    min_length: 20
  }
)

if input[:valid]
  config = input[:processed_input]
  puts "Configuration: #{config}"
else
  puts "Invalid input: #{input[:reason]}"
end

# Get session summary
summary = human_input.session_summary
puts "Human interactions: #{summary[:total_interactions]}"
puts "Approval rate: #{summary[:approvals] / summary[:total_interactions].to_f * 100}%"

Best Practices

1. Strategic Human Checkpoints

Use human approval for irreversible actions (deployments, data modifications)
Request human guidance for strategic decisions
Get human validation for critical outputs

2. Balanced Automation

Enable human input for high-risk operations only
Use timeouts to prevent workflow blocking
Provide meaningful context for human decisions

3. Error Recovery

Always offer human intervention options when tasks fail
Allow humans to modify task parameters for retry
Provide clear error context and suggested actions

4. User Experience

Keep approval requests concise but informative
Provide reasonable timeout values
Show consequences of actions clearly

5. Testing and Development

Use auto-approval mode for testing workflows
Test human interaction flows in development
Monitor interaction patterns and optimize prompts

Configuration Options

# Agent-level human input configuration
agent = RCrewAI::Agent.new(
  name: "agent",
  role: "Specialist",
  goal: "Complete tasks with human collaboration",
  human_input: true,                       # Enable human input
  require_approval_for_tools: true,        # Require approval for all tools
  require_approval_for_final_answer: false # Optional final answer approval
)

# Task-level human input configuration  
task = RCrewAI::Task.new(
  name: "task",
  description: "Complete a complex task",
  agent: agent,
  human_input: true,                       # Enable human input
  require_confirmation: false,             # No confirmation needed to start
  allow_guidance: true,                    # Allow human guidance during execution
  human_review_points: [:completion, :error] # Review at completion and errors
)

# HumanInput utility configuration
human_input = RCrewAI::HumanInput.new(
  timeout: 300,                           # Default timeout in seconds
  verbose: true,                          # Enable verbose logging
  auto_approve: false,                    # Disable auto-approval (for testing)
  approval_keywords: %w[yes y approve],   # Custom approval keywords
  rejection_keywords: %w[no n reject]     # Custom rejection keywords
)

Human-in-the-loop integration makes RCrewAI perfect for scenarios requiring human oversight, strategic input, or collaborative decision-making while maintaining the power and efficiency of AI automation.