Agentic Design Patterns

21 essential patterns for building agentic AI systems. Each one comes with a diagram, a written discussion, a runnable implementation, and a real captured run.

Browse the Patterns · Run the Code · Why the Failing Runs Matter

---

What's Inside

Every pattern is documented four ways, so you can pick whichever gets the idea across fastest.

Format	Location	Good for
Mermaid Diagrams	/mermaid-diagrams	Seeing the flow at a glance, in plain English
Pattern Discussions	/pattern-discussion	When to use it, where it fits, pros, cons, real examples
ASCII Art	/ascii-art	Pasting into Miro, terminals, or anywhere without rich graphics
Runnable Code	/code	Actually seeing it work, plus a saved run for every pattern

Each code example lives in its own folder with the run it produced:

code/
├── resources/                # shared LLM client, output helpers, requirements
└── prompt_chaining/
    ├── code.py               # the pattern, commented
    └── result.txt            # a real captured run, readable without an API key

---

The 21 Patterns

Core Patterns

#	Pattern	What it does	Links
1	Prompt Chaining	Break a complex task into sequential steps	ASCII · Diagram · Discussion · Code · Run
2	Routing	Direct each request to the right handler	ASCII · Diagram · Discussion · Code · Run
3	Parallelization	Run independent work concurrently, then merge	ASCII · Diagram · Discussion · Code · Run
4	Reflection	Critique and revise its own output	ASCII · Diagram · Discussion · Code · Run
5	Tool Use	Call real functions instead of guessing	ASCII · Diagram · Discussion · Code · Run

Advanced Patterns

#	Pattern	What it does	Links
6	Planning	Design the whole step list before executing any of it	ASCII · Diagram · Discussion · Code · Run
7	Multi-Agent Collaboration	Separate roles handing work to each other	ASCII · Diagram · Discussion · Code · Run
8	Memory Management	Decide what to keep verbatim, summarize, or never drop	ASCII · Diagram · Discussion · Code · Run
9	Learning and Adaptation	Turn mistakes into rules, without retraining	ASCII · Diagram · Discussion · Code · Run
10	Model Context Protocol	Reach tools through a contract, not an import	ASCII · Diagram · Discussion · Code · Run

System Patterns

#	Pattern	What it does	Links
11	Goal Setting and Monitoring	Define "done" up front, then measure against it	ASCII · Diagram · Discussion · Code · Run
12	Exception Handling and Recovery	Retry, repair, or fall back, depending on the failure	ASCII · Diagram · Discussion · Code · Run
13	Human-in-the-Loop	Pause for approval on the actions that deserve it	ASCII · Diagram · Discussion · Code · Run
14	Knowledge Retrieval (RAG)	Answer from fetched documents, not from memory	ASCII · Diagram · Discussion · Code · Run
15	Inter-Agent Communication	Address agents by capability over a message bus	ASCII · Diagram · Discussion · Code · Run

Optimization Patterns

#	Pattern	What it does	Links
16	Resource-Aware Optimization	Spend tokens in proportion to the work	ASCII · Diagram · Discussion · Code · Run
17	Reasoning Techniques	CoT, self-consistency, decomposition, and their cost	ASCII · Diagram · Discussion · Code · Run
18	Guardrails / Safety	Screen the input, then screen the output separately	ASCII · Diagram · Discussion · Code · Run
19	Evaluation and Monitoring	Measure quality offline, watch proxies online	ASCII · Diagram · Discussion · Code · Run

Strategic Patterns

#	Pattern	What it does	Links
20	Prioritization	Order work by score, dependencies, and age	ASCII · Diagram · Discussion · Code · Run
21	Exploration and Discovery	Search a space of approaches instead of answering once	ASCII · Diagram · Discussion · Code · Run

---

Running the Code

Examples use LangChain with NVIDIA NIM endpoints.

1. Install from the code/ directory:

cd code
python -m venv .venv
.venv\Scripts\activate                       # Windows
source .venv/bin/activate                    # macOS/Linux
pip install -r resources/requirements.txt

2. Add your key to code/resources/.env:

API_KEY=your_nvidia_api_key

3. Run any pattern as a module, from inside code/:

python -m prompt_chaining.code

Why -m, and why underscored folders?

Running python prompt_chaining/code.py directly fails with ModuleNotFoundError: resources, because Python puts the script's folder on the import path rather than code/. The module form keeps code/ on the path, so every example can share resources/.

That is also why the folders use underscores. python -m prompt-chaining.code is not a valid module name.

To save a run alongside the code, the way every committed result.txt was made:

python -m prompt_chaining.code | Tee-Object prompt_chaining\result.txt

python -m prompt_chaining.code | tee prompt_chaining/result.txt

Piping makes stdout a non-tty, so rich drops colour and falls back to ASCII table borders. The content is identical.

See code/README.md for the full module list and setup detail.

---

About the Saved Runs

Every result.txt is real output from the committed code, not a hand-written illustration. Several of them record the pattern failing, plateauing, or costing more than it saved. Those runs were kept deliberately, because they teach more than a clean one does.

Pattern	What its run actually shows
Evaluation and Monitoring	The scoreboard's verdict is wrong. The grader marked 1 incorrect against an answer key saying one. A broken eval doesn't look broken, it looks like a result.
Reasoning Techniques	All four techniques tie, including the "direct" baseline. Decomposition spent 4 calls to reach what 1 call already had.
Goal Setting and Monitoring	Ends in recorded failure. Fixing the missing specs breaks the word limit, because constraint satisfaction isn't monotonic.
Human-in-the-Loop	The risk assessor gates a routine receipt email as irreversible. Over-gating, live.
Reflection	Never converges. The critic spends three rounds on progressively more marginal objections.
Exploration and Discovery	An earlier run produced no child that beat its parent. Refinement plateaus fast.

Each pattern's discussion doc explains what its run revealed and what to do about it.

Note: temperature is 0.7 and several examples depend on model judgement, so your output will differ in wording and token counts. The structure and conclusions should hold.

---

Source

These patterns are distilled from research on agentic AI systems, made accessible through simple visual representations and clear explanations.

Agentic Design Patterns (full book, PDF)

Contributing

Improvements and additional patterns are welcome. Open an issue or a pull request.

License

MIT, so use these patterns freely in your projects.

_{21 patterns · 21 runnable examples · 21 captured runs}