LLM Agent Architecture Patterns

The Agent Loop

At its core, every LLM agent follows the same pattern:

Observe → Think → Act → Observe → ...

The difference between a chatbot and an agent is autonomy — agents decide what to do next without human intervention at each step.

Pattern 1: ReAct

The simplest effective pattern. The LLM alternates between reasoning and acting:

Thought: I need to find the user's order status
Action: query_database(order_id="12345")
Observation: Order shipped on March 1
Thought: I have the information, I can respond
Action: respond("Your order shipped on March 1")

Pattern 2: Plan-and-Execute

For complex tasks, separate planning from execution:

1. Planner generates a high-level plan
2. Executor handles each step
3. Replanner adjusts when things go wrong

This mirrors how we actually solve problems — think first, then do.

Pattern 3: Multi-Agent

Multiple specialized agents collaborating:

• Researcher gathers information
• Coder writes and tests code
• Reviewer validates output

Choosing the Right Pattern

• Simple Q&A with tools → ReAct
• Multi-step workflows → Plan-and-Execute
• Complex open-ended tasks → Multi-Agent