What Is an AI Agent?

There is no universal definition of AI agent. The older meaning includes any AI system that perceives an environment and acts in it. In current software products, the term usually means a system in which a model directs at least part of the workflow. This page uses that modern, narrower meaning.

The simplest mental model

You give an agent an outcome, not every step.

If you ask ordinary automation to process an invoice, a developer has already written the path: read these fields, apply these rules, update this system. If you ask an agent to resolve an invoice problem, the model may inspect the invoice, look up a purchase order, notice a mismatch, request missing evidence, and then choose whether to update the record or escalate it.

The difference is not that the agent can do anything it wants. The difference is that it has bounded control over what happens next.

A useful test is:

After the system receives a result, who chooses the next step?

If fixed code always chooses, the system is mainly a workflow. If the model can choose a different action based on the result, the system has agent-like control. Real products can mix both: fixed code may handle approvals and payments while a model decides how to investigate an exception.

How an AI agent works

Most current software agents combine an AI model with an AI harness. The model interprets the situation and proposes what to do. The harness supplies context, exposes allowed tools, executes approved actions, records state, and enforces limits.

The basic loop looks like this:

Goal, context, and limits
           |
           v
 Model chooses a next step <------------------+
           |                                  |
           v                                  |
 Harness checks and executes the action       |
           |                                  |
           v                                  |
 Environment returns an observation ----------+
           |
           v
 Finish, ask for help, or stop at a limit

1. The agent receives a goal

The goal describes the desired result. Instructions add constraints such as what may be changed, what requires approval, and what counts as complete. A clear stopping condition matters because “keep trying” is not a safe or useful objective.

2. The model chooses a next step

The model receives the goal and the current state. It might answer, ask a question, select a tool, or decide that it cannot proceed. It can also form or revise a plan, but a detailed plan is not required for every agent.

3. The harness handles the action

A tool is an interface the system can use to retrieve information or change something outside the model. It might search records, run code, read a calendar, or submit an update.

The model normally proposes a tool name and inputs. The harness validates that request, checks permissions, asks for approval when needed, and runs the underlying code. This separation prevents a generated instruction from automatically becoming an authorized action.

4. The environment returns evidence

The tool result becomes an observation: a search result, an error, a test report, a confirmation number, or some other evidence from outside the model. The model uses that observation to choose again.

This feedback closes the loop. Without it, the system may generate a plausible account of success without knowing whether anything actually happened.

5. The loop stops

The agent ends when it reaches a completion condition, cannot continue, needs a human decision, or hits a limit. Limits can include allowed actions, elapsed time, number of steps, or resource use. An agent that can stop safely is better designed than one that merely keeps acting.

A worked example

Imagine asking a delivery agent:

Move Tuesday’s grocery delivery to a time when I am home. Do not change the items.

The agent does not need a hard-coded script for every possible calendar and delivery state.

1. It retrieves the current booking and checks that the requested order is scheduled for Tuesday.
2. It reads your calendar through a tool the harness has permission to use.
3. The observation shows that the current delivery window overlaps an appointment.
4. It requests the delivery service’s available Tuesday windows.
5. It compares those results with the calendar and selects a non-conflicting option.
6. The harness sees that rescheduling changes an external record, so it pauses for your approval.
7. After approval, the harness submits the change.
8. The delivery service returns a confirmation. The agent checks that the day, window, and item list satisfy the original request, then stops.

If no valid window exists, the model might ask whether Wednesday is acceptable. If the calendar tool fails, it might ask you for your availability. If the booking has already shipped, it should explain the blocker rather than invent a successful result.

Those branches reveal the agent. A fixed workflow could perform the same happy-path steps, but the model-directed system can choose a new path from each observation.

Why AI agents matter

Agents are useful when a goal is clear but the required steps vary. They can work through missing information, tool errors, and changing conditions without requiring a person to issue a new instruction after every step.

That flexibility comes with a cost. Each model decision adds time and resource use. A mistaken early action can also shape every later step. Giving a system more tools increases what it can accomplish and what it can damage.

The practical question is therefore not “How autonomous can this be?” It is “Which choices can the system safely make, and where should fixed rules or human approval take over?” Open-ended investigation may tolerate several self-directed steps. Sending money, deleting data, or making a binding decision usually deserves a tighter boundary.

Common misconceptions

“Any app with an AI model is an agent”

A model can summarize a document or answer a message in one call. That is an AI-powered feature, but it is not necessarily an agent. The stronger signal is model-directed control over a process: choosing actions, receiving observations, and deciding what follows.

“One tool call makes it an agent”

A chatbot can always call the same weather function and then display the result. That path may be ordinary automation. A tool becomes part of an agent loop when its result can change what the model decides to do next.

“An agent is just a smarter model”

Model capability helps, but an agent is a system. The model does not create its own database credentials, execute its own code, or grant itself permission. The harness and tool interfaces determine what proposed actions can become real actions.

“Autonomous means unsupervised”

Autonomy is a degree, not an on-off property. An agent can independently research options, pause before a consequential change, and return control when evidence is missing. Checkpoints and escalation are part of the design, not evidence that the system stopped being an agent.

“Agents always learn from experience”

Some agents store information across runs. Others forget everything after a task ends. Memory and learning can expand an agent’s behavior, but goal-directed action does not guarantee either one.

“A confident response means the job is done”

Language models generate likely outputs. Completion is an external fact. A reliable agent checks a receipt, file state, test result, or other observation before it reports success.

How an agent fits into an AI system

The model, harness, and agent are related but not interchangeable.

• The model produces text or structured choices from the context it receives.
• The harness connects the model to instructions, state, tools, permissions, approvals, and logs.
• The agent is the goal-directed system that emerges when the model can use that machinery to direct a process from feedback.

A chatbot may be the interface to an agent, but conversation alone does not make a chatbot agentic. An agent can also run in the background with no chat window at all.

This boundary is partly editorial. Some software frameworks call the configured model component an “agent,” while others reserve the word for the whole running loop. When evaluating a product, ignore the label and inspect the behavior: what goal does it receive, what actions can it take, who chooses each next step, what evidence does it observe, and what makes it stop?

Where to go next

Read What Is an AI Model? to understand the component making the choices. Then read What Is an AI Harness? for the surrounding system that supplies tools, state, permissions, and execution. Continue to What Is Tool Use in LLMs? for the action loop, or use AI Model vs. Chatbot vs. Harness vs. Agent to compare the complete stack.