An Agent Loop is the repeated cycle that lets an AI system work through a multi-step task instead of answering only once.
An Agent Loop is the control cycle that allows an agent to continuously observe, reason, act, and react to new results.
A normal LLM interaction looks like User → LLM → Answer. An agent needs a loop around the model so it can pursue a goal, use tools, inspect what happened, and choose the next step.
Why it matters: An Agent Loop lets an agent carry out a task across multiple steps. Without the loop, the system usually stops after one response.
A normal LLM answers one question once.
An Agent Loop is like an employee who keeps checking the situation, taking one step, checking the result, and choosing the next step until the work is done. This analogy describes a repeated workflow, not human-like thinking.
The loop begins with a goal and ends when the task is complete, a human needs to decide, or a safety limit is reached.
Core sequence: Goal → Observe → Reason → Choose Action → Use Tool → Observe Result → Reason Again → Repeat → Finish
User asks: “Find the API error in the project and fix the code.”
Agent Loop:
This is a loop, not one prompt: the failed test changes what the agent observes, so it reasons again and takes another action.
Goal: understand a support ticket and resolve it.
Observe: read the ticket and account status.
Act: query an approved system or draft a response.
Observe result: check whether the account update worked.
Finish: respond, request approval, or hand the case to a person.
One-shot call: one input produces one output.
Agent Loop: the system can use intermediate results to choose more steps.
Prompt: instructions and context given to a model.
Loop: runtime behavior that repeats model decisions and actions.
Safe loop: has a stop condition, limits, and failure handling.
Infinite loop: keeps running without reaching a controlled end.
These concepts describe different parts of the same agent system:
A system that uses a model, tools, state, and a loop to pursue a goal.
The agent requests a structured action; the application executes the real tool.
Information that can persist or be retrieved across steps or tasks.
Choosing a sequence or next action that moves toward the goal.
The runtime around the model that manages tools, state, permissions, retries, and stops.
Workflow is a defined path; a graph represents possible paths and branches. A loop may be one part of either.
Explore next
An Agent Loop is the repeated cycle of Observe → Reason → Act → Observe Result. It is the foundation that lets an Agent complete multi-step work instead of returning only one answer.
Review the key vocabulary for Agents & Agent Systems.