Building AI Agents with Agentic Patterns
Khan Academy developed Khanmigo using Chain-of-Thought reasoning and ReAct patterns. The agent reasons through math steps, provides hints rather than answers, and achieved a 15% performance improvement across 50+ subjects for 15 million students.
Why is the ReAct pattern particularly suitable for tutoring agents?
Tip: This is a textbook example of ReAct combined with Chain-of-Thought. The agent observes the student's work, reasons through the math steps, then acts by providing targeted hints—mimicking how a human tutor would approach the problem.
JPMorgan deployed COIN (Contract Intelligence) for loan agreement analysis using ReAct with the OpenAI Functions API. The system processes 12,000 contracts annually, saving 360,000 hours of manual work while reducing error rates from 10% to 2%.
How does the OpenAI Functions API benefit contract analysis?
Tip: This demonstrates OpenAI Functions for structured data extraction. The Functions API ensures the agent returns data in predefined schemas, making contract clause extraction reliable and consistent across thousands of documents.
A hospital network deployed LangChain agents for patient triage, reducing wait times by 30% and improving diagnostic accuracy by 18%. The system uses the SQL Database Toolkit with custom prefix/suffix configurations to handle 500+ daily interactions.
What advantage does customizing LangChain agent prefix/suffix provide?
Tip: This demonstrates Customising Standard Agents from Unit 3. By adding healthcare-specific instructions in the prefix/suffix, the agent gains domain expertise that significantly improves triage accuracy and patient safety.