The four core agentic patterns

Reflection

Reflection is the pattern where an agent evaluates and critiques its own output, then uses that critique to improve on the next iteration. The agent is prompted to check its work for correctness, style, and efficiency, and external tools (like unit tests or web searches) can validate results and highlight gaps. In multi-agent setups, reflection can be split across roles, with one agent generating and another critiquing. The survey cites Self-Refine, Reflexion, and CRITIC as methods showing real performance gains from this pattern. It is the most reliable and mature of the four.

Planning

Planning is the pattern where an agent decomposes a complex task into smaller, manageable subtasks, then determines the sequence of steps to accomplish the larger goal. This is essential for multi-hop reasoning and for situations that cannot be scripted in advance. The survey notes a tradeoff: planning enables flexibility, but it produces less predictable outcomes than a fixed, deterministic workflow. You gain adaptability and lose certainty.

Tool use

Tool use is the pattern where an agent interacts with external tools, APIs, or computational resources to go beyond its pre-trained knowledge. It can gather information, run computations, and manipulate data. The survey notes this matured significantly with function calling capabilities in modern models. It also flags a real challenge: when there are many tools available, selecting the right one is hard, and techniques inspired by RAG itself (like heuristic-based tool selection) are used to manage it. This is the same "too many tools causes confusion" problem seen in the context-engineering material.

Multi-agent

Multi-agent collaboration is the pattern where work is split across specialized agents that communicate and share intermediate results, enabling task specialization and parallel processing. Each agent has its own memory and workflow and can itself use tools, reflection, or planning. The survey is honest that this is the least predictable of the four patterns compared to the more mature reflection and tool use, and it points to frameworks like AutoGen, CrewAI, and LangGraph as ways to implement it. This connects directly to the multi-agent topic and its whole debate about reliability.

The survey's framing: these four patterns are the foundation, and agentic RAG systems combine them, from simple sequential steps to adaptive collaborative processes, to handle tasks that exceed traditional RAG.