What are AI agents and how do they work?

What makes software an AI agent? Follow a goal through reasoning, connected tools and permitted actions, with a shopping example and clear limits.

An AI agent is software that pursues a goal using available information and tools it has permission to use. It can choose a next step, inspect what happened and continue until it completes the task or reaches a stopping point.

For a retailer, that distinction matters when software moves from answering a product question to changing a basket or placing an order. The action needs its own permission and a result the store can confirm.

What makes software an AI agent

An AI agent combines a model’s decisions with tools that let it act on a task. The underlying AI model interprets a request and proposes a response or next action. The surrounding software supplies information, executes permitted tool calls and returns their results.

A fixed automated workflow follows steps set in advance, such as sending a dispatch email when an order becomes fulfilled. An agent can choose which available step comes next, such as retrieving a missing product detail before answering a shopper. Anthropic’s explanation of agents and workflows makes this distinction through who controls the sequence.

Products marketed as agents have different levels of independence. Some follow tightly prescribed routes. Others select tools across several steps, with human approval at particular points. A chat interface alone doesn’t tell you which actions the software can perform. The fuller AI agent vs chatbot comparison explains that separate distinction.

“Agentic AI” describes the approach of using AI to direct work towards a goal. An AI agent is a system that applies that approach to a particular job.

From a goal to a completed action

An agent works through a loop of interpreting the goal, taking a permitted step and inspecting the result. A customer request can pass through several tool calls before the agent has enough information to finish.

  1. The customer states the goal. “Find a waterproof coat below my budget” gives the agent a product requirement, but the budget still needs a value.
  2. The agent gathers relevant information. It asks for the missing budget and retrieves catalogue records instead of inventing products from the model’s training.
  3. The model selects a tool and its inputs. A catalogue search can use the entered budget and required product attributes.
  4. The application checks permission and executes the call. A search tool can return matches without having permission to buy them.
  5. The agent reads the result. It can inspect product details, ask a follow-up question or stop if the request can’t be fulfilled.
  6. After an authorised change, the agent checks the resulting store record before reporting success.

Context is the information available for the current decision, including recent messages and tool results. It lets “show me those in navy” refer to coats already discussed. Long-term memory is information saved for retrieval in later conversations, such as a shopping list. The application must provide that storage and retrieval. A model doesn’t automatically remember every earlier session.

A retry can help after a temporary search failure. After an uncertain purchase attempt, the system must first check whether an order exists to avoid placing it twice. A step limit or a request for human help gives the loop an end when it can’t make progress.

A shopping agent at work

shoppa is an AI shopping agent that helps customers find, compare and buy products in one chat on the retailer’s site. It uses the retailer’s catalogue and documents and supports customers after the sale.

Consider a fictional shopper asking for a warm coat under $200. shoppa turns the request into catalogue filters and shows matching products. If the shopper asks for navy, the conversation keeps the coat requirement and budget while refining the results.

The shopper can compare two coats on their recorded material, price and available sizes. Delivery or returns questions draw on the retailer’s documents. Product discovery supplies a choice. It doesn’t authorise payment.

“Add the navy coat in medium” authorises a basket change for that selected item. The resulting basket must show the correct variant, quantity and price. The shopper can then pay in the chat without leaving the retailer’s site. Payment initiation uses the connected checkout and its required authorisation. A completed purchase needs the corresponding payment result and order record, not just a message saying the payment started.

shoppa also supports automatic buying under rules a customer sets in advance. That permission comes from those rules, rather than from a general request to browse. Every shopping agent remains bounded by the products, checkout operations and payment tools its platform supports.

After purchase, the customer can ask shoppa for order status using the order number. When a person needs to take over, shoppa passes the conversation to the retailer’s staff or helpdesk software.

Tools, permissions and human decisions

Tools determine what an agent can request, while access controls determine which requests the application will allow. Reading a product and charging a payment method require different authority.

A useful permission design separates the actions by consequence:

  • Catalogue search reads product information without changing the store.
  • Basket access lets the agent add or remove the shopper’s selected items, with the changed basket visible for inspection.
  • Purchase access commits money and creates an order under an explicit checkout approval or previously configured buying rules.
  • Account access can change details or privileges, so it needs its own identity checks and permission boundaries.

The Model Context Protocol tool specification describes tools with defined inputs and returned results. It also recommends visible tool activity and human control over invocations. Connecting tools is separate from granting unrestricted authority to use them.

A customer’s automatic buying rules might permit one named consumable up to a set total. A substitute outside those rules needs a new decision. The language model can propose an action, but the application and provider must enforce the limits.

Where an agent can fail

An agent can make a plausible recommendation from incomplete information or describe an action that never completed. Check the underlying result at the point where the failure would affect the customer.

Possible failureObservable checkResponse
Stale product informationCompare the selected variant with current price and availabilityRefresh the record before a basket or checkout decision
An unsuitable search resultInspect the recorded attributes against the shopper’s requirementExplain the mismatch and search again or ask the shopper to change a requirement
An ambiguous goalIdentify the missing choice, such as size or whether a budget includes deliveryAsk for that choice before taking the affected action
A failed basket changeRead the basket and confirm the intended item and quantityCorrect the basket before starting payment
An interrupted checkoutCheck payment status and the order recordResume only from the confirmed state or hand over to staff

Incorrect: treating “Your coat is ordered” as proof when the checkout result is missing. Correct: confirming the selected coat in the order record and checking its payment status before saying the purchase succeeded.

Before assigning a store task to an agent, decide what evidence will count as completion. If the agent can’t confirm that result, it must stop or hand the task to someone who can.

Give your store its own shopping agent

Customers ask in their own words. Your agent finds, compares and checks out from your catalogue, right on your site.

See it in actionFree trial · Simple setup