Agentic DesignPatterns and tools for production AI agents

Loading...

📞

Function Calling

Structured interface for AI to invoke external functions and APIs

Complexity: mediumTool Use

In 30 seconds

What: AI invokes external functions via structured schemas, passing validated parameters and receiving typed responses.
When to use: Tasks requiring real-time data, calculations, or state changes that the model cannot perform internally.
Watch out: Models may hallucinate function calls with wrong names, missing parameters, or impossible argument values.

Loading technique guide…

Patterns

Loading...

Loading Agentic Design…

Agentic DesignPatterns and tools for production AI agents

Expert help Pricing

Design Patterns & Techniques

Loading...

📞

Function Calling

Structured interface for AI to invoke external functions and APIs

Complexity: mediumTool Use

In 30 seconds

What: AI invokes external functions via structured schemas, passing validated parameters and receiving typed responses.
When to use: Tasks requiring real-time data, calculations, or state changes that the model cannot perform internally.
Watch out: Models may hallucinate function calls with wrong names, missing parameters, or impossible argument values.

30-Second Overview

Pattern: Enable AI to invoke external functions and APIs through structured schemas

Why: Breaks LLM knowledge limitations, enables real-time data access, external actions, and precise calculations

Key Insight: Schema-driven function calling with parameter validation, parallel execution, and security controls

Quick Implementation

1Define Schema:JSON schema with parameters & types

2Register Tools:Add to function registry with auth

3Validate Params:Sanitize & validate all inputs

4Execute:Call API with timeouts & retries

5Process Results:Normalize & integrate responses

Example: get_weather(location="NYC", units="celsius") → {"temp": 18, "condition": "cloudy"}

Do's & Don'ts

✅Write detailed function descriptions (3-4+ sentences minimum)

✅Use strict JSON schemas with required/optional field validation

✅Implement parameter sanitization and injection attack prevention

✅Add timeouts, retries with exponential backoff, and circuit breakers

✅Use parallel execution for independent function calls

✅Cache deterministic function results for cost optimization

❌Use vague function descriptions that lead to hallucinated parameters

❌Allow unbounded parallel calls without rate limiting

❌Retry non-idempotent operations after timeouts

❌Expose raw API credentials or secrets in function calls

❌Skip input validation and security scanning

When to Use

Use When

• Need real-time data beyond model training cutoff
• Require external API integrations or database queries
• Must perform precise calculations or code execution
• Triggering actions in external systems (email, scheduling)

Avoid When

• Information available in model knowledge with acceptable quality
• Hard real-time requirements where tool overhead breaks SLOs
• High-risk irreversible actions without human approval
• Simple tasks solvable with basic text generation

Key Metrics

Function Selection Accuracy

% correct tool chosen from registry

Parameter Extraction Rate

% required params correctly identified

Schema Validation Success

% calls passing validation checks

API Success Rate

% successful external function calls

Parallel Execution Efficiency

Latency reduction vs sequential

Token Usage Optimization

Context tokens per function call

Security Incident Rate

Injection/XSS attempts blocked

Cost per Task

Total cost: model + API + infrastructure

Top Use Cases

Weather & Real-time Data: get_weather("NYC") → current conditions for decision making

Calendar Integration: schedule_meeting(time, attendees) → automated scheduling workflows

Database Operations: query_db(table, filters) → dynamic data retrieval and updates

Code Execution: run_python(code) → computational tasks and data analysis

Email & Communications: send_email(to, subject, body) → automated notifications

Web Search: search_web(query) → real-time information retrieval

File Operations: read_file(path), write_file(path, content) → document processing

API Integrations: call_api(endpoint, params) → external service interactions

References & Further Reading

Deepen your understanding with these curated resources

Foundational Academic Papers

ReAct: Synergizing Reasoning and Acting in Language Models (ICLR 2023)

Toolformer: Language Models Can Teach Themselves to Use Tools (2023)

Gorilla: Large Language Model Connected with Massive APIs (UC Berkeley 2023)

Berkeley Function Calling Leaderboard (BFCL) - Current Benchmark

Industry Implementation Guides

Anthropic Claude Tool Use Documentation

Google Gemini Function Calling Guide

OpenAI Function Calling Best Practices

LangChain Tools Integration Framework

Enterprise Frameworks & Security

Model Context Protocol (MCP) - Anthropic Standard

Google Agent Developer Kit (ADK) - Enterprise Tools

CrewAI Multi-Agent Tool Coordination

Azure OpenAI Function Calling Security Guidelines

Research & Performance Evaluation

Function Calling in Large Language Models: A Survey (2024)

Tool Use Patterns Overview

API-Bank: Tool-Augmented LLM Benchmark

ToolBench: Tool Learning Evaluation Framework

Loading technique guide…

Patterns

Design Patterns & Techniques

Loading...

30-Second Overview

Pattern: Enable AI to invoke external functions and APIs through structured schemas

Why: Breaks LLM knowledge limitations, enables real-time data access, external actions, and precise calculations

Key Insight: Schema-driven function calling with parameter validation, parallel execution, and security controls

Quick Implementation

1Define Schema:JSON schema with parameters & types

2Register Tools:Add to function registry with auth

3Validate Params:Sanitize & validate all inputs

4Execute:Call API with timeouts & retries

5Process Results:Normalize & integrate responses

Example: get_weather(location="NYC", units="celsius") → {"temp": 18, "condition": "cloudy"}

Do's & Don'ts

✅Write detailed function descriptions (3-4+ sentences minimum)

✅Use strict JSON schemas with required/optional field validation

✅Implement parameter sanitization and injection attack prevention

✅Add timeouts, retries with exponential backoff, and circuit breakers

✅Use parallel execution for independent function calls

✅Cache deterministic function results for cost optimization

❌Use vague function descriptions that lead to hallucinated parameters

❌Allow unbounded parallel calls without rate limiting

❌Retry non-idempotent operations after timeouts

❌Expose raw API credentials or secrets in function calls

❌Skip input validation and security scanning

When to Use

Use When

• Need real-time data beyond model training cutoff
• Require external API integrations or database queries
• Must perform precise calculations or code execution
• Triggering actions in external systems (email, scheduling)

Avoid When

• Information available in model knowledge with acceptable quality
• Hard real-time requirements where tool overhead breaks SLOs
• High-risk irreversible actions without human approval
• Simple tasks solvable with basic text generation