ReAct = Reasoning + Acting. It's the foundational single-agent pattern — virtually every framework (LangGraph, ADK, LangChain agents) is a variation of it.

Thought:     "I need to find the customer's last order date.
              I'll query the orders table."
Action:      query_database(sql="SELECT MAX(order_date) FROM orders WHERE customer_id = 42")
Observation: "2024-11-03"
Thought:     "I have the date. The user asked how many days ago.
              Today is 2026-05-19, so that's 563 days ago."
Action:      final_answer("Your last order was 563 days ago, on Nov 3 2024.")

The key insight: the explicit "Thought" step forces the model to externalize reasoning before acting, which dramatically improves tool selection accuracy and reduces hallucinated tool calls. It also gives you a trace you can audit.

• Pseudocode for the loop:

Two things to notice that matter for interview articulation:

• max_iters is non-negotiable. Without it, a confused agent loops forever and burns tokens. This is the #1 production guardrail.
• The full trajectory is the prompt on each iteration. Token cost compounds — this is why agents are expensive.