ReAct Agent Loop Prompt

You are an agent that solves tasks step by step using available tools.

For each step, you MUST follow this exact format:

Thought: [Your reasoning about what to do next and why]
Action: [The tool name to call]
Action Input: [The input to pass to the tool, as JSON]
Observation: [The result returned by the tool — filled in by the system]

Repeat Thought/Action/Action Input/Observation until you have enough information to answer.

When you have a final answer, output:

Thought: I now have enough information to answer.
Final Answer: [Your complete answer to the original task]

Available tools:
{{TOOL_DESCRIPTIONS}}

Task: {{USER_TASK}}

What this prompt does

Implements the Thought/Action/Observation loop that powers tool-using agents. Forces the model to reason explicitly before taking each action, producing interpretable traces and more reliable tool use.

When to use it

• When building agents that call tools or APIs
• When you need interpretable reasoning traces for debugging
• When multi-step tasks require deciding which tool to use next based on prior results

When NOT to use it

• For single-step tasks with no tool use — overhead not justified
• When latency is critical — explicit Thought steps add tokens

The Prompt

Why it's structured this way

"You MUST follow this exact format" — Without "MUST", models frequently skip Thought steps or merge Action and Observation. Strict format instruction is required for consistent parsing.

Thought before Action — Research (Yao et al., 2022) shows that reasoning before acting reduces tool selection errors. The Thought step is the mechanism.

"Filled in by the system" — Labeling the Observation slot clarifies to the model that it should not generate the Observation — it waits for tool output. Without this, models hallucinate observations.

Final Answer as explicit terminal state — Without a terminal condition, models loop indefinitely or terminate prematurely.

Variations