Complete setup instructions for testing the new ReAct (Reasoning + Acting) agent

Version: 1.1-beta
Branch: feature/one-react-agent

ℹ️ Note: This guide is for the experimental ReAct agent - a single AI agent that thinks and acts like an engineer, reading docs, writing code, testing, and debugging iteratively. This is different from Tako's multi-agent orchestration system.

• What is the ReAct Agent?
• Installation
  • Prerequisites
  • Windows Installation
  • Linux/macOS Installation
• Running the ReAct Agent
  • Basic Usage
  • Example Queries
  • Understanding the Output
• Performance Expectations

The ReAct (Reasoning + Acting) agent is a single AI agent that operates differently from Tako's traditional multi-agent system:

1. 🤔 Reasons about your query and plans an approach
2. 📚 Reads Okta API documentation to understand endpoints
3. 💻 Writes Python code to answer your question
4. 🧪 Tests the code with sample data (LIMIT 3)
5. 🐛 Debugs failures by analyzing errors and retrying
6. ✅ Iterates until the code works correctly
7. 📝 Generates production-ready script with full data

• Acts like a person: Thinks through problems step-by-step, debugs itself
• Works with smaller models: gemini-2.0-flash, claude-3.5-haiku, gpt-4o-mini
• Self-correcting: Sees errors, understands them, and fixes the code
• Production-ready output: Generates CSV files you can share

• Use ReAct for: Complex queries requiring API exploration, certificate analysis, or custom reporting
• Use Multi-Agent for: Standard queries, web UI experience, hybrid SQL + API mode

Before you begin installation, ensure you have the following:

• 🐍 Python 3.9+ - ReAct agent requires Python 3.9 or newer
• 🏢 Okta tenant with read-only admin access - You'll need API token with read permissions
• 🤖 Access to an AI provider - You'll need at least one of the following:
  • Google Gemini (Vertex AI) - Recommended: Fast & Cheap
  • Anthropic Claude
  • OpenAI
  • Azure OpenAI

⚠️ Note: The ReAct agent uses significantly more tokens (~200k-400k per query) compared to the standard multi-agent system. We recommend using cheaper models like gemini-2.0-flash-exp or claude-3.5-haiku.

git clone -b feature/one-react-agent https://github.com/fctr-id/okta-ai-agent
cd okta-ai-agent

# Create virtual environment
python -m venv venv

# Activate virtual environment
.\venv\Scripts\Activate.ps1

# If you get an execution policy error, run this first:
# Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser

# Install requirements
pip install -r requirements.txt

# Copy the sample environment file
Copy-Item .env.sample .env

Edit the .env file with your settings:

Required Configuration:

# 🔐 Okta Configuration
OKTA_CLIENT_ORGURL=https://your-okta-domain.okta.com
OKTA_API_TOKEN=your_okta_api_token_here

# 🧠 AI Provider (choose ONE)
AI_PROVIDER=vertex_ai  # or anthropic, openai, azure

# 🌐 Vertex AI Configuration (Recommended - Fast & Cheap)
VERTEX_AI_SERVICE_ACCOUNT_FILE=C:\path\to\your\service-account.json
VERTEX_AI_REASONING_MODEL=gemini-2.0-flash-exp
VERTEX_AI_CODING_MODEL=gemini-2.0-flash-exp

Alternative AI Providers:

Anthropic Claude (Fast & Affordable):

AI_PROVIDER=anthropic
ANTHROPIC_API_KEY=your_anthropic_api_key_here
ANTHROPIC_REASONING_MODEL=claude-3-5-haiku-20241022
ANTHROPIC_CODING_MODEL=claude-3-5-haiku-20241022

OpenAI (Slower but works):

AI_PROVIDER=openai
OPENAI_API_KEY=your_openai_api_key_here
OPENAI_REASONING_MODEL=gpt-4o-mini
OPENAI_CODING_MODEL=gpt-4o-mini

Optional Circuit Breaker Configuration:

# Prevents infinite loops (defaults are fine for testing)
REACT_AGENT_MAX_SQL_EXECUTIONS=5
REACT_AGENT_MAX_API_EXECUTIONS=20
REACT_AGENT_MAX_SCHEMA_LOADS=3
REACT_AGENT_MAX_ENDPOINT_FILTERS=10

git clone -b feature/one-react-agent https://github.com/fctr-id/okta-ai-agent
cd okta-ai-agent

# Create virtual environment
python3 -m venv venv

# Activate virtual environment
source venv/bin/activate

# Install requirements
pip install -r requirements.txt

# Copy the sample environment file
cp .env.sample .env

Edit the .env file with the same configuration as Windows (see above), but use Unix-style paths:

VERTEX_AI_SERVICE_ACCOUNT_FILE=/path/to/your/service-account.json

The ReAct agent runs as a command-line tool that generates Python scripts based on your queries.

Windows:

python scripts\okta_react_agent_test.py "Your query here"

Linux/macOS:

python scripts/okta_react_agent_test.py "Your query here"

# Simple user query
python scripts\okta_react_agent_test.py "List all active users"

# Application and certificate analysis
python scripts\okta_react_agent_test.py "Find all SAML applications and their certificate expiry dates"

# Multi-entity query
python scripts\okta_react_agent_test.py "List users who logged in during the last 10 days and fetch all their assigned applications and group memberships"

# Certificate compliance
python scripts\okta_react_agent_test.py "Find all apps with active certificates, get the certificate expiry dates, and identify which ones expire in the next 90 days"

The ReAct agent will show its reasoning process:

🤔 Reasoning about query...
📚 Reading Okta API documentation...
💻 Writing test code...
🧪 Testing with sample data (LIMIT 3)...
✅ Test passed! Generating production code...

You'll see output like:

💾 Generated code saved to: c:\...\okta-ai-agent\src\core\data\okta_query_20251021_143022.py

▶️  To execute the script, run:
   python c:\...\okta-ai-agent\src\core\data\okta_query_20251021_143022.py

📝 TEST SUMMARY
================================================================================
Query: Find all SAML applications and their certificate expiry dates
Status: ✅ SUCCESS
Time taken: 142.3s
Token usage: 236,124 input / 2,850 output / 238,974 total

Run the generated script to get your actual results:

Windows:

python src\core\data\okta_query_20251021_143022.py

Linux/macOS:

python src/core/data/okta_query_20251021_143022.py

This will create a CSV file:

Results saved to: c:\...\okta-ai-agent\src\core\data\results_20251021_143025.csv
Total records: 47

Open the CSV file in Excel, Google Sheets, or any text editor:

Windows:

# Open in default CSV viewer
Invoke-Item src\core\data\results_20251021_143025.csv

# Or view in Excel
Start-Process excel.exe src\core\data\results_20251021_143025.csv

# Or view in Notepad
notepad src\core\data\results_20251021_143025.csv

Linux/macOS:

# View with default CSV viewer
open src/core/data/results_20251021_143025.csv  # macOS
xdg-open src/core/data/results_20251021_143025.csv  # Linux

# Or view in terminal
cat src/core/data/results_20251021_143025.csv

• Simple queries (1-2 API calls): 30-60 seconds
• Medium queries (3-5 API calls): 1-2 minutes
• Complex queries (6-10 API calls): 2-5 minutes
• Very complex queries (10+ API calls): 5-8 minutes

• Simple queries: 100k-200k tokens (~$0.05-0.20)
• Medium queries: 200k-300k tokens (~$0.20-0.40)
• Complex queries: 300k-500k tokens (~$0.40-1.00)

Costs assume gemini-2.0-flash-exp ($0.075/$0.30 per 1M tokens). OpenAI costs 10-20x higher.

• Generated Python script: src/core/data/okta_query_TIMESTAMP.py
• CSV results: src/core/data/results_TIMESTAMP.csv
• Both files include timestamps for easy tracking

Found a bug? Open an issue: https://github.com/fctr-id/okta-ai-agent/issues

Need help? Reach out to [email protected] or Slack me in the JAMF/Okta community

Want to contribute? PRs welcome on the feature/one-react-agent branch!

Have a query that failed? Share the query and error logs - it helps improve the agent!

🎉 You're All Set! The ReAct agent is now ready to explore your Okta environment with its reasoning and self-debugging capabilities. Start with simple queries and gradually explore more complex scenarios as you see how the agent thinks through problems.

🌟 © 2025 Fctr. All rights reserved. Made with ❤️ for the Okta community.

Not affiliated with Okta, Inc.