What is an AI agent?

In the language-model setting, an AI agent is a software system in which a model helps choose the next steps toward a task and coordinates available tools. This definition focuses on how the system operates, not a promise that it will act correctly or finish without supervision.

Anthropic’s agent architecture guide distinguishes predefined workflows from agents in which a language model dynamically directs the process and tool use. The guide also acknowledges that “agent” is used for different designs.

Decide which choices belong to the agent

The useful question is which decisions you want the software to make. Selecting the next document to inspect is different from deciding to send a message or change an account. Specify the task, the available information and the actions that need approval before evaluating the result.

The model, tool and feedback loop

Start with the task and available state

Give the system an objective and the information it is permitted to use. For a support draft, that might include the customer’s question and an approved help collection. Record what is known, what is missing and what a successful draft must contain.

Choose the next permitted step

The model can propose a next action using the task and earlier results. In Anthropic’s architecture description, tool results provide feedback for assessing progress, and an agent can pause at a checkpoint or when blocked.

Software executes the tool request

Claude’s tool-use documentation separates client tools, which your application executes and returns results from, from server tools executed on Anthropic’s infrastructure. A model’s request and the operation that software performs are separate parts of that interface.

For your own system, enforce access and approval checks where the operation runs. Giving a tool a reassuring description does not establish that its handler is restricted to the intended files, accounts or actions.

Saved information needs an explicit scope

Claude’s documented memory tool stores and retrieves information in files you control. That describes persisted information, not a guarantee that an agent remembers everything or improves after every task.

Decide what state is saved, who can read it and when it is discarded. Reusing a previous result is different from training a model; inspect the implemented mechanism rather than assuming either from the word “memory.”

Describe access and supervision

Compare actual systems by the tasks they can attempt, their read and write access, and the approval they require. A research-only configuration and one permitted to change records have different review needs, even when they use the same model. A product label does not answer those questions.

A hypothetical support-draft task

Imagine a fictional support agent asked to draft an answer about a product setting. It may search an approved help collection at most twice. It has no tools to send messages, issue refunds or change accounts. The required output is a draft with supporting references, or a clear escalation if the evidence is missing.

Illustrative support-draft run and its controls
SituationProposed next stepController checkAcceptable result
The question names a product settingSearch the approved help collectionRead access only; stay within the search limitRelevant documentation, or a recorded lack of evidence
A search fails or evidence conflictsInspect the failure and decide whether another search is justifiedNo unlimited retry; stop when the limit is reachedExplain the unresolved point and escalate
The draft is readyPresent it for reviewNo sending or account-change tool is availableA referenced draft, not a claim that a message was sent

This is a design example, not a deployed service or a measured result. A real implementation would need to enforce these limits and verify that the draft’s references support its answer. Stopping without an answer can be the correct outcome when the permitted evidence is insufficient.

A chat interface can contain an agent

Anthropic describes customer-support agents that combine a familiar chatbot interface with tool integration. A chat window therefore does not establish whether a system only generates text or also takes actions. Inspect its tools and execution controls.

Inspect behaviour behind the label

“Bot,” “assistant” and “agent” are product labels as well as technical terms. Check whether the system follows fixed steps, lets a model choose steps, or combines the two. Avoid using the name alone to infer access, reliability or independence.

Name the system and the approach

Use AI agent for the concrete system being inspected. Agentic AI describes the broader approach to goal-directed behaviour and workflow autonomy. Neither phrase specifies a particular permission set or level of human supervision.

Test completion, failure and stopping

Anthropic’s architecture guide recommends stopping conditions such as a maximum number of iterations and warns about compounding errors. Completion, missing evidence, denied access and exhausted limits should each have an explicit path.

In Anthropic’s agent-evaluation terminology, a task has defined inputs and success criteria; a transcript records what happened, while an outcome is the final state of the environment. Inspect the claimed result as well as the steps that produced it.

For the support example, check that no message was sent, that the references match the setting, and that an unavailable help collection produces an escalation. Include these failure cases when testing a changed model, prompt or tool handler.

Define the human handoff

Evaluate a proposed task separately from a whole job. Name who reviews unresolved questions and who is responsible for approving external actions. A demonstration of one completed draft does not establish that every responsibility in a role can be delegated.

Budget for attempts and tool work

Claude’s tool-use documentation lists input tokens, output tokens and applicable server-tool charges as cost components. Those are that platform’s billing inputs, not a universal price per completed task.

Set a run budget that accounts for retries and tool work, then inspect actual usage alongside whether the task was completed. Do not infer a fixed task cost from a single model response or substitute an arbitrary trial period for evaluation.

Sources and scope

The cited documents explain Anthropic’s architecture, Claude’s tool execution and agent-evaluation terminology. The architecture article notes that its tooling landscape has changed since publication; no framework catalogue or performance result is used here. The support task is hypothetical, and these design checks do not certify a deployment as safe.