What is an AI agent?
An AI agent uses a model to choose steps and use permitted tools toward a goal. It can look up information and prepare work, with limits on what it may do and when a person must approve.
By Exemplary AI · Updated · Illustrated explainer
The key idea
An agent combines a goal, information and tools. Permission to act must be defined separately.
The idea, illustrated.
The agent loop, with a place for people
Figure 03Check a purchase request against policy.
Look up the request and the relevant rule.
Is a required detail still missing?
Ask for the detail or draft the review.
A spending decision stays with the authorised person.
- 01
Read the goal
The agent receives a task, such as checking a purchase request against a budget and policy.
- 02
Use a permitted tool
It selects an available action, examines the returned information and chooses what to do next.
- 03
Finish or ask for help
It returns a result, requests missing information or stops at the approval point defined for the task.
What makes this different from a chat response?
A text-only reply can describe how to check a purchase request. An agent connected to the relevant systems can look up the request, read the policy and prepare a review. Tools are the specific actions an application exposes, such as searching a folder or reading a record.
Anthropic’s engineering guidance distinguishes a workflow with predefined paths from an agent that chooses its own sequence of actions. Everyday products use the word “agent” more broadly, so the useful question is what the system actually does, what it can access and where it stops.
Primary source: Anthropic: Building effective agents
Good agents have clear boundaries
A goal like “handle procurement” leaves too much unspecified. A narrower instruction can say which records to inspect, what output to produce and which actions require review. The application must enforce access; a sentence in a prompt is not a substitute for permission checks.
Tool design also affects reliability. An action with a clear purpose, required inputs and understandable result is easier to use correctly. A timeout should be treated as an uncertain outcome, not proof that an action failed. Before retrying an action that changes a record, the system needs a way to check what happened.
Primary source: Anthropic: Writing effective tools for AI agents
An everyday example
Preparing a purchase review
The agent reads the request, finds the applicable policy and lists missing details. It prepares a draft recommendation with links to the records it used.
An authorised person makes the spending decision. The agent reduces the information-gathering work while the approval rule remains explicit.
A common misconception
“An agent needs access to everything to be useful.”
A well-scoped agent needs only the information and actions required for its task. More access can create unnecessary exposure and harder-to-understand behaviour.
How Exemplary helps
You define the job. We wire up the AI.
Exemplary brings instructions, models, knowledge and permitted tools together. We help configure the task and its review points so the agent fits the way your team works.
Explore agents in ExemplarySources & further reading
Primary publications and documentation. Accessed 22 September 2026.
- 01Building effective agents
Anthropic · Engineering article · 2024
- 02Writing effective tools for AI agents
Anthropic · Engineering article · 2025