AI Agent Patterns for bootcamp students. Each pattern is a single, focused file demonstrating core agentic AI concepts. Provider-agnostic - works with 100+ LLM providers via litellm (Ollama, OpenAI, Gemini, Claude, Mistral, and more). Switch providers with a single environment variable - no code changes needed!
These patterns are just the beginning. If you want to master production-ready AI engineering and build real applications that users love, check out my AI Bootcamp for Software Engineers.
Use code GITHUB300 for €300 off the AI Bootcamp for Software Engineers!
Next Cohort: November 3rd, 2025
Enroll Now with GITHUB300 Discount
This repository contains the foundational patterns you'll master in the bootcamp. Ready to go from demos to production? Join the next cohort!
# 1. Setup
make setup
# 2. Configure LLM provider (or use Ollama locally - no API key needed)
cp env.example .env
# Edit .env and set LITELLM_MODEL (e.g., 'ollama/llama3.2' or 'openai/gpt-4')
# Add API keys for cloud providers if needed
# 3. Run all patterns
make run
# 4. Or run individual patterns
python patterns/01_prompt_chaining.py
python patterns/02_routing.py
python patterns/03_parallelization.py- 01_prompt_chaining.py - Chain prompts together
- Automated RFP writer that extracts requirements from client brief → summarizes key needs → generates full proposal draft with cover letter and deliverables
- FAQ summarizer that extracts answers from docs → reformats → writes human-friendly version
- 02_routing.py - Route requests to different handlers
- Customer service triage that routes incoming messages to "refunds", "shipping", or "technical" LLM sub-agents for specialized responses
- In-app chatbot that detects sentiment → routes angry users to human escalation pipeline
- 03_parallelization.py - Run multiple LLM calls in parallel
- SEO content pipeline that generates meta title, meta description, blog intro, and social post captions simultaneously
- Resume enhancer that parallel-generates versions optimized for LinkedIn, ATS, and portfolio
- 04_reflection.py - Use reflection to improve code/output quality
- Autonomous code reviewer that generates code → reviews it for style/performance → regenerates improved version
- Blog summarizer that re-evaluates tone ("too formal?" → "rewrite in friendlier tone")
- 05_tool_calling.py - Use tools with LLMs (Native Function Calling)
- Smart operations assistant that can fetch Jira tickets, query database metrics, and trigger Slack updates via tool APIs
- Calendar assistant that uses a "schedule" tool to book meetings from natural text
- Real-world example: REST Countries API integration for country information
- 06_planning.py - Break down complex tasks into plans
- Marketing campaign planner that creates multi-step plan with timelines and assets needed for "launch new product" goal
- Travel planner that turns "3 days in Rome" into structured itinerary
- 07_multi_agent.py - Coordinate multiple AI agents
- AI newsroom with reporter agent drafting story → editor agent refining tone → fact-checker agent verifying sources
- Video creation bot with separate agents for script, voiceover, and thumbnail
- 08_memory_management.py - Manage conversation memory and context
- AI career coach that remembers user's goals, prior feedback, and performance over time to offer context-aware coaching
- Customer chatbot that remembers user preferences (size, brand, last order)
- 10_mcp.py - Use external tools via Model Context Protocol (Real MCP Servers)
- Enterprise assistant that interacts with CRM, internal wiki, and file servers securely via standardized protocol
- AI that can query Notion, Jira, and Google Drive seamlessly using MCP connectors
- Real examples: Filesystem MCP server, Git MCP server
- 11_goal_setting.py - Set goals and monitor progress
- Sales enablement AI that sets quarterly sales targets, tracks progress, and suggests next best actions
- Habit tracker AI that checks daily progress and adjusts reminders dynamically
- 12_exception_handling.py - Handle exceptions and recover gracefully
- Financial automation bot that handles failed API calls or invalid data by retrying, alerting, or falling back to cached data
- Chatbot that detects "API unavailable" and apologizes with an alternate response
- 13_human_in_loop.py - Integrate human oversight and intervention
- Legal AI that drafts contracts and flags ambiguous clauses for lawyer review
- Social media moderator AI that sends uncertain posts to a human for manual approval
- 14_knowledge_retrieval.py - Retrieve and use domain knowledge
- Customer success AI that answers queries using company-specific knowledge base + recent case studies
- AI that looks up specific product specs or FAQ answers before replying
- 15_inter_agent_communication.py - Enable communication between agents
- Smart factory system where maintenance agent reports sensor issues → logistics agent orders parts → scheduler agent delays affected lines
- Fitness coach AI coordinating with nutrition agent to align meal and workout plans
- 16_resource_optimization.py - Optimize compute/resources
- Multi-tenant AI SaaS that dynamically adjusts LLM model size based on user tier or query complexity
- Chatbot that uses GPT-4 for "creative writing" and GPT-3.5 for "FAQ lookup"
- 17a_chain_of_thought.py - Show reasoning steps
- AI math grader that solves student problem step-by-step and checks each reasoning step
- Expense classifier showing reasoning for why a transaction is "travel" vs "meals"
- 17b_self_correction.py - Self-validation and repair
- Codegen AI that generates script → runs unit test → debugs based on output automatically
- Text summarizer that re-checks if summary misses any key entity
- 17c_problem_decomposition.py - Break complex problems into parts
- Startup advisor AI that breaks "launch a product" into market research, MVP, and growth phases
- Homework helper that splits a multi-step word problem before solving
- 18_guardrails.py - Implement content/policy filters
- AI writing assistant that filters confidential data, PII, or bias-inducing terms before output
- Prevents profanity or misinformation in AI-generated tweets
- 19a_evaluation.py - Evaluate LLM output
- AI tutor QA system that grades AI explanations based on clarity, correctness, and engagement
- Automated prompt evaluator scoring outputs for fluency and factuality
- 19b_monitoring.py - Monitor agent performance
- AI customer support dashboard that tracks response latency, satisfaction scores, and escalation rates
- Alerts when AI's error rate exceeds threshold or response time spikes
- 20_prioritization.py - Manage task queues
- AI ops scheduler that prioritizes background jobs based on urgency, dependency, and ROI
- Email assistant that prioritizes unread emails by importance (CEO > newsletter)
- 21_exploration_discovery.py - Explore unknowns
- R&D assistant that finds unexplored research areas and relevant citations automatically
- News summarizer that surfaces trending but underreported stories
- 22_pydantic_validation.py - Data schema validation
- LLM data ingestion pipeline that validates structured outputs (JSON, schema-bound fields) before storing
- API wrapper that ensures AI responses fit a strict ProductSchema
- 23_agentic_rag.py - Agentic retrieval-augmented generation
- Corporate knowledge AI that fetches from internal docs, emails, and dashboards, then synthesizes actionable insight
- AI that answers "What did we discuss last meeting?" by retrieving meeting notes
- 24_workflow_orchestration.py - Orchestrate complex flows
- Automated video production system with script writer → voiceover → visual generation → upload
- Job application AI that drafts resume → tailors cover letter → applies automatically
- 25_subgraphs.py - Modular AI components
- AI architecture library with reusable subgraphs for parsing input, enriching context, and output formatting
- Plug-in LLM component that handles summarization across different workflows
- 26_state_machines.py - Finite state workflow logic
- Loan processing system that transitions between "application submitted → under review → approved → funded"
- Conversational flow that tracks user from greeting → query → solution → feedback
- 27_recursive_agents.py - Self-referencing reasoning
- Research planner AI that recursively delegates subtasks like literature review → summarization → synthesis
- Debugging AI that recursively investigates root causes in logs
- 28_code_execution.py - Safely run generated code
- Data analyst copilot that generates code to compute KPIs, runs safely in sandbox, and returns results
- Spreadsheet assistant that writes & executes formulas based on user queries
- 29_query_rewriter.py - Rewrite and optimize queries for better retrieval
- Search engine that rewrites user queries to improve results: "how to code" → "Python programming tutorial for beginners"
- Customer support that expands vague questions into specific, actionable queries
- 30_relevancy_check.py - Check and filter relevant information
- Content moderation system that scores articles for relevance to user interests
- Research assistant that filters academic papers by relevance to research topic
- 31_data_processing.py - Process and structure data for RAG systems
- Document processor that cleans, chunks, and structures PDFs for knowledge base
- Data pipeline that transforms raw text into searchable, categorized content
- 32_plan_executor.py - Execute structured plans with monitoring
- Project management AI that breaks down tasks into steps and tracks progress
- Workflow automation that executes complex business processes step-by-step
- 33_anonymization.py - Anonymize and de-anonymize data for privacy
- Healthcare system that removes patient names while preserving medical data
- Customer data processor that anonymizes PII for analytics while maintaining reversibility
- 34_react.py - Reason + Act Loop (ReAct)
- Customer support AI that reasons through issues step-by-step, queries database for answers, and takes live actions (e.g., refund, escalate, notify user)
- Data-cleaning agent that identifies problem entries, applies fixes, then verifies results dynamically
- 35_pev.py - Plan–Execute–Verify (PEV)
- Financial trading bot that plans trades, executes safely, and verifies profit/loss outcomes before next cycle
- DevOps AI that deploys updates → tests functionality → rolls back automatically if verification fails
- 36_blackboard.py - Collaborative Multi-Agent Blackboard
- Research assistant system where multiple domain agents (e.g., legal, financial, tech) post and refine shared insights in a central memory board
- Bug-triage AI where analysis, fix, and QA agents collaborate on one shared state to converge on a solution
- 37_episodic_semantic_memory.py - Dual Memory: Episodic + Semantic
- AI tutor that recalls what topics a student struggled with (episodic) while leveraging structured knowledge (semantic) for tailored lessons
- Personal assistant that remembers past interactions and learns factual context like birthdays, preferences, or goals
- 38_tree_of_thoughts.py - Tree of Thoughts (ToT) Reasoning
- Strategic planner that explores multiple reasoning paths for a problem (e.g., marketing strategies) and picks the most optimal route
- Problem-solver that branches reasoning alternatives, evaluates partial outcomes, and merges best insights
- 39_mental_loop.py - Internal Simulation Loop
- Robot controller that simulates movements or actions internally before committing to real-world execution
- Financial decision engine that runs "mental" dry-runs of portfolio changes before actual trades
- 40_meta_controller.py - Meta-Agent Task Router
- AI manager that routes incoming requests to best-fit specialist agents (e.g., "analytics", "design", "sales")
- Multi-department enterprise assistant that orchestrates and supervises agents across different workflows
- 41_graph_memory.py - Graph-Structured World Model
- Research summarizer that extracts entities and relations (people, projects, papers) into a knowledge graph
- Customer insight AI that links interactions, purchases, and feedback in a connected data graph for richer reasoning
- 42_ensemble.py - Ensemble of Experts
- Content generator that combines outputs from multiple expert AIs (tone, grammar, SEO) before final synthesis
- Decision engine that aggregates recommendations from several models and selects the most consistent result
- 43_dry_run.py - Dry-Run Simulation Harness
- Financial AI that simulates trades or investment decisions and requires approval before live execution
- Operations assistant that runs workflow simulations to validate downstream safety before deployment
- 44_rlhf.py - Reinforcement Learning from Human Feedback
- Content generation system that continuously improves responses based on human ratings and feedback loops
- Customer interaction AI that refines tone and accuracy using user satisfaction scores
- 45_cellular_automata.py - Local Interaction → Global Behavior
- Traffic optimizer where each vehicle agent reacts locally (speed, distance) to produce smooth flow system-wide
- Urban planning simulator where thousands of micro-agents self-organize housing, transport, and energy networks
- 46_reflexive_metacognitive.py - Self-Aware and Self-Evaluating Agent
- Medical assistant AI that measures its confidence per diagnosis and automatically escalates uncertain cases to human experts
- Customer service bot that knows when it's unsure or stuck and requests clarification or human takeover
All patterns use litellm as a unified abstraction layer. Switch providers with a single environment variable - no code changes needed!
-
Ollama (Recommended for local testing - No API key needed):
# Install Ollama: https://ollama.ai ollama pull llama3.2 ollama serve # In .env: LITELLM_MODEL=ollama/llama3.2
-
OpenAI:
# In .env: LITELLM_MODEL=openai/gpt-4 OPENAI_API_KEY=your_key_here -
Google Gemini (Free tier available):
# In .env: LITELLM_MODEL=gemini/gemini-1.5-flash GOOGLE_API_KEY=your_key_here -
Other Providers:
# Mistral LITELLM_MODEL=mistral/mistral-small MISTRAL_API_KEY=your_key_here # Anthropic Claude LITELLM_MODEL=anthropic/claude-3-sonnet-20240229 ANTHROPIC_API_KEY=your_key_here # Fireworks AI LITELLM_MODEL=fireworks/accounts/fireworks/models/llama-v3p1-8b-instruct FIREWORKS_API_KEY=your_key_here
To switch providers, just change LITELLM_MODEL in your .env file:
LITELLM_MODEL=ollama/llama3.2→ Use OllamaLITELLM_MODEL=openai/gpt-4→ Use OpenAILITELLM_MODEL=gemini/gemini-1.5-flash→ Use Gemini
No code changes needed! All patterns work identically across all providers.
See env.example for complete configuration options.
- Prompt Chaining - Break complex tasks into steps
- Routing - Smart delegation based on request type
- Parallelization - Speed up multiple LLM calls
- Reflection - Improve code quality through self-review
- Tool Calling - Use external tools with LLMs
- Planning - Break down complex tasks into structured plans
- Multi-Agent - Coordinate multiple specialized agents
- Memory Management - Manage conversation context and user preferences
- MCP - Connect to external tools via Model Context Protocol
- Goal Setting - Set and track progress toward objectives
- Exception Handling - Handle errors and recover gracefully
- Human-in-the-Loop - Integrate human oversight and intervention
- Knowledge Retrieval - Use RAG for better responses
- Inter-Agent Communication - Enable agents to work together
- Resource Optimization - Optimize for cost vs performance
17a. Chain-of-Thought - Step-by-step reasoning process 17b. Self-Correction - Review and improve solutions 17c. Problem Decomposition - Break down complex problems
- Guardrails - Implement safety and content filtering 19a. Evaluation - Assess AI response quality and safety 19b. Monitoring - Track performance and behavior metrics
- Prioritization - Manage tasks and workload intelligently
- Exploration & Discovery - Research and knowledge discovery
- Pydantic Validation - Type-safe data validation
- Agentic RAG - Intelligent retrieval-augmented generation
- Workflow Orchestration - Coordinate complex AI workflows
- Subgraphs - Create modular, reusable components
- State Machines - Implement event-driven agent behavior
- Recursive Agents - Build self-referencing AI systems
- Code Execution - Safely execute code with AI agents
- Query Rewriter - Rewrite and optimize queries for better retrieval
- Relevancy Check - Check and filter relevant information
- Data Processing - Process and structure data for RAG systems
- Plan Executor - Execute structured plans with monitoring
- Anonymization - Anonymize and de-anonymize data for privacy
- ReAct - Dynamic reasoning and action loops
- PEV - Plan-Execute-Verify workflows
- Blackboard Systems - Multi-agent collaboration
- Episodic + Semantic Memory - Dual-memory systems
- Tree of Thoughts - Systematic reasoning exploration
- Mental Loop - Simulation before execution
- Meta-Controller - Intelligent task routing
- Graph Memory - Structured knowledge storage
- Ensemble - Multiple perspective analysis
- Dry-Run Harness - Safety-critical execution
- RLHF - Self-improvement through feedback
- Cellular Automata - Emergent behavior systems
- Reflexive Metacognitive - Self-aware AI systems
All providers work identically - same code, same interface, just change LITELLM_MODEL:
- Ollama - Local models (no API key needed) - Recommended for testing
- OpenAI - GPT-4, GPT-3.5-turbo, and more
- Google Gemini - Free tier available (gemini-1.5-flash, gemini-1.5-pro)
- Anthropic Claude - Claude 3 Sonnet, Opus, and more
- Mistral - European models (mistral-small, mistral-large)
- Fireworks AI - Fast inference models
- And 100+ more - See litellm providers
Provider-agnostic design: Switch providers with one environment variable. No code changes needed!
Each pattern is a single, focused file - perfect for bootcamp students:
- ✅ One concept per file - No cognitive overload
- ✅ Self-contained - Each pattern runs independently
- ✅ Simple & clear - Easy to understand and modify
- ✅ Progressive complexity - Build from basics to advanced
- ✅ Real examples - Working code you can run immediately
- ✅ Provider-agnostic - Works with any LLM via litellm
- ✅ Real APIs - REST Countries API, MCP servers, and more
Patterns 05 (Tool Calling) and 34 (ReAct) demonstrate native function calling using litellm's unified interface. Works identically across all providers:
- Define functions with JSON schemas
- LLM automatically chooses when to call functions
- Execute functions and inject results
- Works with Ollama, OpenAI, Gemini, and more
Example: patterns/05_tool_calling.py uses REST Countries API for real-world function calling.
Pattern 10 demonstrates real MCP server integration:
- Filesystem MCP Server: List, read, write files
- Git MCP Server: Check git status, view commit logs
- Standardized protocol for external tool integration
Example: patterns/10_mcp.py shows real MCP server usage (not mocks).
- Bootcamp Students - Progressive learning with single-file patterns
- AI Developers - Quick reference for common agentic patterns
- Educators - Ready-to-use examples for teaching
- Researchers - Foundation for building complex AI systems
- Start with Foundation - Run patterns 01-08 first
- Explore Advanced - Try patterns 10-16 for complex behaviors
- Master Reasoning - Practice patterns 17a-17c for better AI thinking
- Build Systems - Combine patterns 23-28 for production systems
- Master RAG - Learn patterns 29-33 for advanced retrieval systems
- Advanced Architectures - Explore patterns 34-46 for cutting-edge agentic systems
These patterns are just the foundation. To build production-ready AI systems that solve real business problems, join the AI Bootcamp for Software Engineers.
- 6 weeks of hands-on learning with live sessions
- Production-ready AI projects you'll build and deploy
- Expert support from Param Harrison (15+ years engineering experience)
- Real applications like voice assistants, research agents, and automated workflows
- Production focus - not just demos, but systems that scale
Use code GITHUB300 for €300 off!
Next Cohort: November 3rd, 2025
Enroll Now | Download Syllabus
Perfect for bootcamp students learning AI agent development. Ready to build production systems? Join the next cohort!