An agent in artificial intelligence is a system that receives information about its environment and selects actions to pursue a goal.
The word agent has been used in artificial intelligence for decades. It does not refer only to modern tools powered by large language models. A chess program, a robot, and an AI coding assistant can all be described as agents, even though they work in very different ways.
Today, people often use “AI agent” to describe a software system that uses a large language model, tools, and a multi-step process to complete a task. That is one modern kind of AI agent—not the only meaning of the term.
A simple example of an AI agent
Imagine a system asked to compare three project-management products.
It might:
- Receive the goal: compare the products using their features and pricing.
- Look for relevant information, such as product pages and pricing pages.
- Notice when a piece of information is missing.
- Search for another source or mark the information as unavailable.
- Organize the results into a comparison table.
- Stop when it has enough information or when it cannot verify a detail.
This is a teaching example, not a claim that every AI agent can browse websites or verify prices. An agent can only use the information and tools it has access to. If it cannot access a source, it should say so rather than imply that it checked.
How does an AI agent work?
At a high level, an AI agent connects observations, decisions, and actions:
Observe the situation → choose an action → see what happens → decide what to do next
The product-comparison example shows how the parts of an agent fit together:
- An environment: the situation it operates in—in this case, the product pages and other sources it can access.
- Observations: information it receives, such as a listed feature, a price, or a page it cannot open.
- A goal or objective: what it is trying to accomplish, such as comparing three products by their features and pricing.
- Actions: the steps it can take, such as opening an available source, searching for missing information, or adding a finding to the comparison table.
- A decision process: how it chooses the next step, such as checking another source when a product's pricing is unclear.
- A stopping condition: when it should finish, ask for help, or report a limitation—for example, when the comparison is complete or a missing detail cannot be verified with the available sources.
Suppose one product page says “contact sales” instead of listing a price. That observation affects the next decision: the agent could check another available official source or mark the price as unavailable. If it checks another source, the result becomes a new observation. The process continues until it can produce the comparison or explain what remains unknown. It should not invent a price to fill the gap.
A traditional agent may follow rules or use a decision-making algorithm. A modern AI agent may use a language model to interpret instructions, choose among available tools, and work through several steps.
The details depend on the system. Some agents act autonomously; others require approval before important actions. Some use tools; others only select an answer or action based on their input.
Common types of agents in AI
AI textbooks often describe several classic agent types. These categories are useful for understanding different approaches, though real systems can combine ideas from more than one category.
1. Simple reflex agents
A simple reflex agent chooses an action based on what it observes right now.
For example, a basic thermostat might switch on the heating when the measured temperature falls below a set threshold. It does not need to build a detailed history of the room to apply that rule.
Useful when: the situation is simple and the right action can be chosen from the current observation.
Limitation: it may struggle when a decision depends on information that is not visible right now.
2. Model-based reflex agents
A model-based reflex agent keeps an internal representation of the situation. That can help it make decisions when it cannot observe everything directly.
For example, a robot vacuum may use its recent observations to keep track of areas it has already visited.
Useful when: the system needs some memory of how the environment changes.
Limitation: its decisions depend on whether that internal model is accurate enough.
3. Goal-based agents
A goal-based agent considers a desired outcome and chooses actions that may help reach it.
For example, a navigation system can compare possible routes against the goal of reaching a destination.
Useful when: there may be several actions or routes that could achieve the goal.
Limitation: knowing the goal does not always tell the agent which option is best when there are trade-offs.
4. Utility-based agents
A utility-based agent evaluates options using a measure of how desirable each outcome is. It can help when there is more than one important consideration, such as time, cost, or risk.
For example, a route planner might weigh travel time against tolls or distance.
Useful when: the system needs to compare different outcomes rather than simply reach a goal.
Limitation: the result depends on how the system represents and weighs those trade-offs.
5. Learning agents
A learning agent improves its behavior based on data or feedback.
For example, a recommendation system may adjust what it suggests after observing which recommendations users interact with.
Useful when: experience or feedback can help the system make better decisions over time.
Limitation: learning from data does not guarantee that the system will behave correctly in every new situation.
These categories describe broad approaches, not a checklist that every modern AI product fits into exactly.
Traditional AI agents vs. LLM-based agents
The term agent is broader than today’s large language model products.
| Traditional AI agent | LLM-based agent |
|---|---|
| May use rules, search, planning, or learned policies | Uses a large language model as part of its decision process |
| Often operates in a narrowly defined environment | May interpret open-ended language instructions |
| May use structured inputs and actions | May work with text, code, tools, or other connected systems |
| Can be designed without a language model | Usually relies on a language model, but may also use rules and conventional software |
A large language model by itself is not necessarily an agent. It may simply generate a response to a prompt. A system that uses a model as part of a goal-directed process—with a way to observe results and take actions—may be described as an LLM-based agent.
The label alone does not tell you how much the system can do. Check its actual tools, permissions, and behavior.
AI agents vs. chatbots
A chatbot is an interface for communicating with software. An AI agent describes a system’s role in pursuing a goal. The two ideas can overlap: an AI agent may be accessed through a chat interface.
A basic chatbot may answer a question in one response. An agent may be able to work through a sequence of steps, use tools, and check the results—if those capabilities are available and enabled.
| Question | What to check |
|---|---|
| Can it use tools? | Which tools can it access, and what can it do with them? |
| Can it take actions? | Does it only suggest an action, or can it carry it out? |
| Does it keep working across steps? | Can it observe results and decide what to do next? |
| Does it need approval? | Which actions require a person to review or confirm them? |
A chat interface alone does not make a system an agent.
A modern example: an AI coding assistant
Suppose you ask a coding assistant to review a change for possible accessibility issues.
A carefully designed process could be:
- Identify which changed files are in scope.
- Read the project’s accessibility guidance.
- Inspect the relevant code.
- Report possible issues with an explanation of the evidence.
- Stop and ask before making changes, if the task was review-only.
- Leave the final decision to a person who can inspect the code and run the project’s checks.
The assistant might use project instructions or a reusable skill to follow the same review process consistently. The skill provides task-specific guidance; it is not necessarily the agent itself.
If you use Claude Code, explore the Claude Code Skills directory to find reusable guidance for a coding workflow. To understand how skills are structured, read What Are Agent Skills? and What Is SKILL.md?.
What can AI agents do?
Depending on how they are built and what they can access, AI agents may help with tasks such as:
- Finding and organizing information.
- Following a repeatable workflow.
- Reading or editing files.
- Calling APIs or interacting with connected software.
- Comparing options against stated requirements.
- Producing a report, draft, or other output.
- Checking whether a task appears complete.
These are possible uses, not guaranteed features. An agent cannot reliably perform an action unless it has the necessary access, instructions, and tools.
What are the limitations and risks?
AI agents can misunderstand instructions, use incomplete information, or produce an incorrect result. If they can act on files or external systems, a mistake may have consequences beyond a wrong answer.
Use safeguards that match the risk of the task:
- Give the system only the permissions it needs.
- Review important actions before they are carried out.
- Ask it to identify the sources or evidence behind important claims.
- Test a workflow on a low-risk example first.
- Keep human review for sensitive, irreversible, or high-impact actions.
- For coding work, review changes and run the project’s tests before merging.
A system’s confidence is not evidence that its output is correct. When an agent cannot verify something, it should make that limitation clear.
How to get started with AI agents
Start with a small task that has a clear outcome and is easy to check.
- Choose one task. Pick something specific, such as summarizing a set of documents or reviewing a code change.
- Define what a good result looks like. State the expected output and any boundaries.
- Check the system’s capabilities. Confirm what information and tools it can access.
- Try a low-risk example. Do not begin by granting broad permissions.
- Review the result. Check whether it used the right information and followed the instructions.
- Make useful instructions reusable. If the process works, consider using a suitable skill, template, or workflow.
If you are exploring coding workflows, start with a relevant resource in the Skills directory. Choose one that matches your task and tool rather than installing a large collection without a clear use case.
Frequently asked questions
What is an agent in artificial intelligence?
An agent in artificial intelligence is a system that observes information about an environment and takes actions to achieve a goal. The term includes many kinds of systems, not only those powered by large language models.
What is an artificial intelligence agent?
An artificial intelligence agent is a system that uses AI to choose or perform actions in pursuit of an objective. Its design may use rules, planning, machine learning, a language model, or a combination of approaches.
Is an AI agent the same as an AI model?
No. An AI model processes input and produces output. An AI agent is a larger system that may use a model as part of a process involving goals, observations, decisions, and actions.
Is an AI agent the same as a chatbot?
No. A chatbot is a way to interact with software. An agent is a system designed to pursue a goal. A chatbot can be the interface to an agent, but not every chatbot performs multi-step actions.
Is an AI agent always autonomous?
No. Some systems act with limited or no approval during a task; others require a person to confirm actions. Autonomy depends on the system’s design, permissions, and settings.
Is a skill the same as an AI agent?
No. In coding tools, a skill is generally reusable guidance for a task. An AI agent is the system pursuing a goal. An agent may use a skill, but the skill is not necessarily an agent by itself.
What is the difference between an agent and a workflow?
A workflow is a process made up of steps. Those steps may be fixed or may include decisions made by an AI system. An agent may use a workflow, but the terms do not mean the same thing.
What should I use first: an agent, a workflow, or a skill?
Use the simplest option that suits the task. A fixed workflow can work well when the steps are predictable. A skill can make task-specific guidance reusable. An agent may be useful when the system needs to respond to changing information or choose among actions.
Key takeaway
An AI agent is a system that observes information and takes actions toward a goal. Traditional AI agents and modern LLM-based agents can work in different ways, so evaluate the system by its actual capabilities, tools, permissions, and safeguards—not by the word “agent” in its product description.