The 21 agentic design patterns

The foundational design pattern where a complex task is decomposed into a linear sequence of smaller, discrete LLM calls.

Dynamically directing a user request to the most appropriate specialized agent, model, or processing path based on the semantic intent and.

Executing multiple independent agentic tasks simultaneously — such as voting on a decision, generating multiple creative drafts, or verifying facts.

Enables an agent to critique its own output (or the output of another agent) to identify errors, hallucinations, or areas for improvement, and then.

Extends the capabilities of an LLM by allowing it to interact with the external world (APIs, databases, calculators, software environments) to.

The capability of an agent to break down a high-level, ambiguous user goal into a sequence of executable steps (a plan), and then execute them while.

Structuring a system as a team of specialized agents (e.g., a Researcher, a Writer, a Reviewer, a Coder) that collaborate, hand off tasks, and.

Mechanisms for agents to store, index, and retrieve information over time, spanning short-term (conversation context) and long-term (knowledge.

The ability of an agent to improve its performance over time based on feedback, user interactions, or new data, without full model retraining.

A standardized protocol (Model Context Protocol) for connecting AI models to data sources and tools, ensuring consistent context and state exchange.

Mechanisms for agents to define high-level objectives, break them into trackable sub-goals, and continuously monitor progress towards completion.

Strategies for agents to detect failures (e.g., API errors, hallucinations, logic loops) and self-correct or retry using alternative strategies.

Integrating human oversight at critical decision points to validate agent actions, provide feedback, or disambiguate requests.

Retrieval-Augmented Generation (RAG) equips agents with the ability to query external knowledge bases (Vector DBs, Search Indices) to ground their.

Protocols and formats for autonomous agents to exchange messages, tasks, and state information to collaborate on a shared goal.

Agents aware of their token consumption, API costs, and computational limits, optimizing their strategies accordingly (e.g., using a cheaper model.

Advanced cognitive architectures like Chain-of-Thought (CoT), Tree-of-Thought (ToT), and ReAct (Reason+Act) that structure the model's internal.

Architectural safeguards (input/output filters) to prevent agents from executing harmful actions, leaking PII, or deviating from policy.

Frameworks for measuring agent performance (accuracy, faithfulness, tool usage) and monitoring behavior in production (traces, logs).

The ability of an agent to rank disparate tasks or goals based on urgency, importance, and constraints, effectively managing its own backlog.

Agents proactively seeking new information, testing hypotheses, or exploring an environment to expand their knowledge or capabilities, rather than.