Agentic Design Patterns
A pattern catalog for architecting LLM-powered autonomous systems — agents that reason, use tools, remember, and self-correct. Framework-agnostic, model-agnostic, production-tested.
Most LLM applications are request-response: prompt in, text out. Agentic systems break that model — they pursue goals over multiple steps, use tools, maintain state, and correct their own mistakes. The gap between a working demo and a production system is where most teams get stuck. These patterns bridge that gap.
Pick a starting point
Building your first agent? Start with a Single Agent, add Chain of Thought reasoning, and connect tools with a Tool Router.
Scaling to multi-agent? Compare Orchestrator vs Pipeline vs Swarm, then add Shared Memory for coordination.
Improving reliability? Layer Guardrails on output, add Reflection for self-correction, and build an Eval Suite for regression testing.
Debugging failures? Check the Anti-Patterns — you're probably hitting the Prompt Monolith, Infinite Loop, or Amnesiac Agent.
The catalog
Agents
How agents are structured: scope, authority, and coordination model.
Single Agent · Role-Based · Hierarchical · Swarm
Memory
How agents remember: buffers, summaries, vector stores, knowledge graphs, and cross-session persistence.
Buffer · Summarization · Vector Store · Extracted Facts · File-Based · Knowledge Graph · Shared Memory
Orchestration
How work flows: pipelines, routing, parallelism, and human checkpoints.
Orchestrator · Pipeline · Parallel · Router · Evaluator-Optimizer · Human-in-the-Loop
Reasoning
How agents think: chain of thought, action loops, upfront planning, and self-critique.
Chain of Thought · ReAct · Plan-and-Execute · Reflection
Tools
How agents use the world: function calling, retrieval, code execution, and tool discovery.
Tool Router · RAG · Code Execution · MCP
Output
How agents deliver: schemas, streaming, safety validation, and source attribution.
Structured Output · Streaming · Guardrails · Citation
Evaluation
How you measure: LLM judges, test suites, domain metrics, and human feedback loops.
LLM-as-Judge · Test-Driven · Domain Metrics · Eval Suite · Human Feedback
Anti-Patterns
What to avoid: the mistakes that break agents in production.
Framework Trap · Prompt Monolith · Infinite Loop · Vibe Deployment · Agent Sprawl · Tool Junk Drawer · Happy Path Mirage · Amnesiac Agent
Real World
How production systems are actually built. Architecture breakdowns mapped to the patterns in this catalog.
Claude Code · ChatGPT · Cursor · Devin · GitHub Copilot · Perplexity · OpenClaw