Gartner LLM-based AI Design Patterns
Overview
Emerging Patterns for Building LLM-Based AI Agents
Patterns for Building LLM-Based Agents
Gartner's diagram outlines patterns across three categories: Agent Architectures, Functional Patterns, and Operational Patterns for LLM-based AI agents.
Agent Architectures
- Solo Agent: Atomic, monolithic agents for simple tasks.
- Agent Roles: Define personas to guide behavior and scope.
- Agent-to-Agent Handoff: Delegation in multi-agent systems.
- Multi-Agent Modularity: Decomposes tasks across specialized sub-agents.
Functional Patterns
- Prescribed Plan: Fixed workflows for repeatable processes.
- Dynamic Plan: Runtime strategy generation for adaptability.
- MHQA: Multi-hop question answering with iterative retrieval.
- Collaborating Agents: Dynamic teamwork.
- Orchestrated Agents: Predefined graphs for controlled interactions.
- LLM Interaction Patterns: ReAct (reason-act loops), Reflexion (self-critique), Chain of Thought (step-by-step reasoning).
Operational Patterns
- Agent Evaluation: User-in-the-loop oversight; LLM-as-judge (deterministic).
- Agent Action Patterns: Function calling, generated code, API tools.
- Memory Patterns: RAG (retrieval-augmented generation), memory longevity, memory scope.
- Security Patterns: LLM guardrails, identity tokens, logging.