AI Agents and Autonomy
Most people's first experience with modern AI is a chatbot: you type a message, the AI replies, and the conversation ends there. But a growing class of AI systems does much more. They receive a goal, form a plan, execute a series of actions across multiple steps, check the results, adjust the plan, and keep going until the goal is reached — or until they get stuck. These systems are called AI agents, and they represent a qualitative shift in how AI can be used.
What Is an AI Agent?
An AI agent is a system that perceives its environment, decides on actions, takes those actions, observes the results, and uses the results to decide what to do next. The key word is autonomy: the agent is not just answering a question — it is pursuing a goal through a sequence of decisions. A simple example: you give an agent the goal of booking a flight, a hotel, and a car for a business trip. The agent searches flight options, picks the best one based on your preferences, fills in the booking form, confirms the reservation, then moves on to the hotel, and then the car — all without you needing to supervise each step. If the first hotel choice is sold out, it adapts and tries the next option.
A chatbot responds to a single message with a single response and waits. An AI agent receives a goal and takes a sequence of actions — possibly over minutes or hours — to achieve it. Agents can make decisions, call tools, check results, and iterate, often without requiring human input at each step.
How Agents Work: The Perceive-Plan-Act Loop
Most AI agent architectures are built around a loop. The agent perceives the current state of the world — reading a file, looking at a web page, checking the results of a previous action. It plans the next step, often by prompting a language model to reason about the situation. It acts by calling a tool or taking a real-world action. It then observes the result and loops back to perceiving the new state. This loop can run hundreds of times for a single complex task. The agent might draft a document, check it against a rubric, revise it, check again, and repeat — all autonomously. Longer loops create more capable agents, but they also mean errors can compound: a mistaken assumption at step 3 may not become visible until step 47, by which time the agent has done significant work in the wrong direction.
Types of Autonomy
Agents differ in how much autonomy they exercise. A low-autonomy agent checks with the human user at every decision point. A high-autonomy agent only interrupts the user at the beginning and end, handling everything in between on its own. More autonomy means faster task completion and less cognitive load on the human — but also less opportunity to catch mistakes. The appropriate level of autonomy depends on how high the stakes are and how much the agent's judgment can be trusted for that specific task. Booking a restaurant reservation is low-stakes; autonomously sending legal documents is not.
The longer an agent runs without human review, the more a small early mistake can snowball. In safety research, this is called error compounding or cascading errors. Human checkpoints — moments where a person reviews the agent's progress before it continues — are a key safeguard for high-stakes agentic tasks.
Match each agentic AI concept to its correct description.
Terms
Definitions
Drag terms onto their definitions, or click a term then click a definition to match.
What fundamentally distinguishes an AI agent from a chatbot?
Why might high autonomy be dangerous for an AI agent working on an important task?
Design an Agent with Checkpoints
- Step 1: Imagine you are building an AI agent to help a student research and draft a five-paragraph essay on a topic of their choice.
- Step 2: List every action the agent would need to take, in order, from receiving the assignment to delivering the finished draft.
- Step 3: Identify at least three checkpoints — specific moments where a human should review the agent's work before it continues.
- Step 4: Explain why you placed a checkpoint at each of those moments. What could go wrong if the agent continued without review?
- Step 5: Write a sentence describing what level of autonomy you would give this agent and why.