Agentic Design Patterns

21 runnable reference implementations for understanding common LLM-agent workflow patterns.

This repository is a clean-room educational reference library containing runnable implementations of 21 agent design primitives. Each directory isolates one pattern and pairs the implementation with a small practical example.

The goal is not to present one universal agent framework. It is to make individual orchestration ideas easy to inspect, run, compare, and combine.

Project status: educational reference implementations. CI verifies repository structure and Python syntax; examples that call external model APIs are not treated as production systems or exercised end to end without credentials.

Patterns

#	Pattern	Reference use case
1	Prompt Chaining	Sequential document processing
2	Routing	Request classification and delegation
3	Parallelization	Concurrent analysis and synthesis
4	Reflection	Generate, critique, refine
5	Tool Use	Function calling and external tools
6	Planning	Goal decomposition and task execution
7	Multi-Agent	Specialized agent collaboration
8	Memory	Conversation state and persistence
9	Adaptation	Feedback-driven behavior changes
10	MCP	Standardized tool/resource integration
11	Goal Setting	Goal breakdown and progress tracking
12	Exception Handling	Fallback and recovery workflows
13	Human-in-the-Loop	Escalation and human feedback
14	RAG	Retrieval-augmented generation
15	Inter-Agent Communication	Agent-to-agent messaging
16	Resource Optimization	Caching, batching, and cost control
17	Reasoning	Structured problem-solving workflows
18	Guardrails	Validation and policy checks
19	Evaluation	LLM-as-a-Judge and quality assessment
20	Prioritization	Priority-based task ordering
21	Exploration	Testing multiple candidate approaches

Quick start

Clone the repository:

git clone https://github.com/josephsenior/Agentic-Design-Patterns.git
cd Agentic-Design-Patterns
cp .env.example .env

Move into a pattern directory, install its dependencies, and run the example:

cd prompt_chaining
pip install -r requirements.txt
python example.py

Some examples use Gemini. Add the required API key to the pattern's .env file when needed.

Repository philosophy

Each implementation is intentionally small and pattern-focused:

pattern_name/
├── implementation.py
├── example.py
├── README.md
├── requirements.txt
└── .env.example

Exact filenames can vary by pattern, but the examples are organized so the orchestration primitive remains easy to identify.

How the patterns relate

Sequential and planning

• Prompt Chaining
• Planning
• Goal Setting
• Prioritization

These patterns organize work over time and make task progression explicit.

Decision and feedback

• Routing
• Reflection
• Adaptation
• Exploration

These patterns select, critique, revise, or compare candidate paths.

Tools and coordination

• Tool Use
• Multi-Agent
• Inter-Agent Communication
• MCP

These patterns connect agents to capabilities or to other specialized agents.

State and knowledge

• Memory
• RAG

These patterns provide persistent conversational state or external knowledge retrieval.

Reliability and control

• Exception Handling
• Human-in-the-Loop
• Guardrails
• Evaluation
• Resource Optimization

These patterns help make agent workflows more observable, bounded, and testable.

Example combinations

Combination	Use case
Routing + Tool Use	Select the right capability for each request
Planning + Evaluation	Validate progress before advancing
RAG + Tool Use	Combine retrieved context with external actions
Memory + Adaptation	Use historical feedback to adjust behavior
Multi-Agent + Inter-Agent Communication	Coordinate specialized roles
Guardrails + Exception Handling	Bound actions and recover from failures

Who this is for

• Engineers learning agent orchestration beyond single-prompt applications
• Students comparing common LLM-agent patterns
• Builders looking for small references before integrating a pattern into a larger system

For a larger autonomous-agent runtime that combines planning, tools, execution, durability, and validation, see Grinta.

Contributing

Issues and pull requests are welcome.

• Open an issue
• View the repository

License

MIT — see LICENSE.