What is an AI agent?
An AI agent is an application that pursues a goal across multiple steps: a model interprets the task, chooses or requests tool calls, receives results, keeps relevant state, and continues, stops or asks for approval. A chatbot can answer a turn; a workflow follows fixed code paths; an agent has some model-directed control. Its permissions and guardrails—not the label—define what it can actually do.
Why — the first-principles explanation
A language model by itself produces a response from an input. It does not automatically have access to your inbox, database, calendar or payment system. An AI agent is the surrounding application that gives a model a bounded job, instructions, tools and a way to continue across steps.
The core cycle is simple: interpret the goal, choose a next action, request a tool call, let the application execute that call, inspect the result, update state and decide whether to continue, stop or ask a person. The model does not magically execute a function; the application decides which requested calls are valid, runs them with its credentials and can reject them. That boundary is crucial for security and billing.
OpenAI describes agents as applications that plan, call tools, collaborate across specialists and keep enough state to complete multi-step work. Anthropic uses a useful architectural distinction: a workflow runs model and tool steps through predefined code paths, while an agent lets the model dynamically direct parts of its process and tool use. Google distinguishes agents from assistants and bots by degree of autonomy, complexity and interaction. A real product can combine all three patterns, so the label alone is not a technical specification.
An agent can be useful for open-ended work such as investigating a support request, checking several systems, preparing a draft or repairing code after tests fail. It is a poor fit when a deterministic rule or short workflow is cheaper, easier to audit and just as effective. More loops mean more latency, model/tool cost and opportunities for an early mistake to affect later steps. OpenAI’s documentation treats guardrails, human review, approvals, state, tracing and evaluations as separate production concerns.
The practical definition is therefore bounded autonomy. An agent may choose a next step, but its tools, credentials, data scope, budget, stop conditions and approval gates constrain the result. Before trusting one, ask what it can read, what it can change, which actions require confirmation, how runs are logged and evaluated, and what happens when a tool fails or the agent cannot finish.
An example that makes it click
Imagine a support assistant with access to an order system and a policy database. It can find an order, check delivery status, draft a response and propose a refund. A rule can make all refunds read-only drafts, while a human must approve the final payment action. The model chooses the sequence, but the application’s permissions and approval gate decide what actually happens.
How to do it
- Define the goal and the allowed side effects. Reading data, drafting a message, changing a record and sending money need different controls.
- List the tools, credentials and data the agent may use. Start with the narrowest permissions and read-only access where possible.
- Specify state and stop conditions: what the agent may remember, how long it can run, how much it can spend and when it must escalate.
- Add validation, policy checks and human approval before irreversible or high-impact actions. Make tool failures and missing information explicit.
- Evaluate representative tasks, adversarial inputs, tool errors and incomplete data. Track success, factuality, unwanted actions, latency, cost and human handoffs.
- Explain the boundary to users: which decisions the model can suggest, which code enforces, what data is retained and how a person can review or undo an action.
Key facts
- OpenAI defines agents as applications that plan, call tools, collaborate across specialists and keep enough state to complete multi-step work.
- Anthropic distinguishes workflows with predefined code paths from agents in which the language model dynamically directs its process and tool use.
- Anthropic recommends using agents when the required steps are hard to predict and notes the trade-off of higher cost and compounding errors.
- Google distinguishes AI agents, assistants and bots by autonomy, complexity and interaction; an assistant can act under user supervision, while a bot commonly follows predefined rules.
- An agent’s model requests a tool call, but the surrounding application executes it and can enforce permissions, validation, approval and logging.
- OpenAI documents guardrails, human review, resumable approvals, state, tracing and evaluations as separate parts of building reliable agent workflows.
- An AI agent is not automatically conscious, generally intelligent, accurate, secure or authorized to act. Those properties depend on the model, application design, data and controls.
Choose an agent by its boundaries
Translate the definition into a work use case, a cost model and a tool comparison you can verify.
▶ The 60-second explainer (script)
What is an AI agent? It is an application that pursues a goal across multiple steps. A model interprets the task, requests a tool call, receives the result, updates its state and decides whether to continue, stop or ask for approval. The application—not the model by itself—executes the tool and enforces permissions. A chatbot usually answers a turn. A workflow follows paths written in advance. An agent has some model-directed control over the path. That flexibility helps with open-ended support, research or coding tasks, but it adds latency, cost and error risk. OpenAI documents guardrails, human review, approvals, tracing and evaluations as production concerns. So ask what the agent can read, change and remember, who approves risky actions, and how failures are handled. The word agent is less important than the boundary around it.
What authoritative sources say
People also ask
What is the difference between an AI agent and a chatbot?
A chatbot can answer a turn. An agent can continue through a bounded loop: it calls an allowed tool, reads the result, chooses another step and stops or requests approval. A product can contain both a chat interface and agentic runs.
What is the difference between an AI agent and a workflow?
A workflow follows predefined code paths. An agent lets the model dynamically choose parts of the process and tool use. Production systems often combine them, using deterministic checks around a model-directed section.
Does an AI agent have memory?
Only when the application stores and retrieves state. That can be conversation history, a database row, files or a task record. Ask what is retained, for how long, who can access it and how it can be deleted.
Can an AI agent send emails or make payments?
It can request those actions only if the application exposes the tools and credentials. A safe design limits permissions and requires confirmation or human approval before sending messages, changing records, spending money or deleting data.
Are AI agents safe?
Safety depends on the task and controls, not the name. Use least privilege, validation, guardrails, budgets, logging, evaluation, clear stop conditions and human review for consequential actions. No model removes the need for application security.
Can an AI agent work offline?
Yes, if the model and tools run locally or on an internal network. A cloud model or external service requires connectivity, and local execution still needs permissions, data controls and a way to update and evaluate the system.
How much does an AI agent cost?
There is no universal price. Cost can include model calls, tool/API usage, storage, hosting, monitoring, failed retries and human review. Compare cost per completed task and failure, not only a model’s per-token price.
The same question, asked other ways
- What are AI agents?
- What is an agent in AI?
- What is AI agent?