Agentic patterns define the control flow: how the agent decides when to think, when to act, and when to stop. ReAct (Reason + Act) alternates: the model outputs a "Thought" (reasoning), then an "Action" (e.g. a tool call), then the app returns an "Observation" (tool result); this repeats until the model outputs an "Answer". Plan-and-execute first asks the model to produce a full plan (step 1, 2, 3…), then the app runs each step and feeds results back. Tool-augmented (single-call) gives the model access to tools in one round; it may return one or more tool calls, and you can run them and optionally do one more LLM call with the results. Choosing a pattern affects latency, cost, and how well the agent handles multi-step tasks.
Common agentic patterns
- 🔄
ReAct
Reason + Act in a loop (thought → action → observation)
- 📋
Plan-and-Execute
Plan steps first, then execute each step
- 🔧
Tool-augmented
Single call with tools; model chooses which to use
How an agent runs: the loop
The agent keeps looping (think → act → observe) until it produces a final answer or hits a step limit.
ReAct in practice: one loop
Reason (Thought) → Act (call a tool) → Observe (get result). Repeat until the agent can answer.
Example: Plan-and-execute in practice
Task: "Compare the top 3 laptops on this site." Plan: (1) Fetch the product page, (2) Extract names and prices, (3) Summarize. The app executes each step, passes results to the model, and the model produces the final comparison.