5 Agentic AI Design Patterns

Agentic behaviors allow LLMs to refine their output by incorporating self-evaluation, planning, and collaboration. Five named patterns cover most of what 'agentic' means in practice:

1. Reflection — the agent reviews and self-critiques its own output, then iterates to improve it, rather than accepting its first attempt as final. This is the conceptual seed behind the Loop multi-agent pattern (see 7 Patterns in Multi-Agent Systems) and behind automated prompt/output optimizers like Opik's MetaPromptOptimizer (see Automated Agent/Prompt Optimization with Opik) — both are, at their core, systematized reflection.

2. Tool Use — giving the agent access to external functions/APIs it can call to extend its capabilities beyond pure text generation. This is covered in depth in this topic's own architecture/workflow content above, and in a full hands-on build (native CrewAI tool + MCP server) in Building Custom Tools for Agents.

3. ReAct (Reason + Act) — explicitly combines Reflection-adjacent reasoning ('Thought') with Tool Use ('Action'), interleaving the two in a loop until enough information is gathered to answer. Covered in full hands-on depth, including a framework-free implementation, in ReAct Implementation From Scratch.

4. Planning — decomposing a larger task into a sequence of smaller steps before execution begins, rather than trying to solve everything in one shot. Framework grounding: CrewAI supports this via a planning=True flag on a Crew, which has an LLM generate a step-by-step plan before any agent starts executing tasks.

5. Multi-Agent — delegating parts of a task to multiple specialized agents that collaborate, rather than relying on one agent to do everything. Covered in full depth, including 7 distinct orchestration topologies, in 7 Patterns in Multi-Agent Systems.

In practice, these patterns compose: a production agentic system commonly combines Planning (decompose the task) with Multi-Agent (delegate the pieces) with ReAct (each sub-agent reasons and acts) with Reflection (each sub-agent or a review-agent critiques and iterates on outputs) with Tool Use (sub-agents call real APIs).

# CrewAI's planning=True flag — a concrete example of the Planning pattern
from crewai import Crew, Process

crew = Crew(
    agents=[researcher, writer, reviewer],
    tasks=[research_task, writing_task, review_task],
    process=Process.sequential,
    planning=True,  # an LLM generates a step-by-step plan before execution begins
)
result = crew.kickoff()