What is an AI agent?

In ecommerce, "agent" gets used for almost everything: the chatbot in the corner of the site, an automation that sends emails, the sales rep who answers the phone. Here an AI agent is something more precise: a program that takes an assignment and decides for itself the steps to complete it, using tools such as searching a catalog or creating a cart. It matters to a store because that program already visits stores on shoppers' behalf, and what it finds depends on what the store lets it read and do.

How it works

OpenAI defines it in one sentence: an agent is a system that completes tasks on your behalf with a high degree of independence. Its guide to building them asks for three traits. The agent uses an LLM (large language model) to direct execution and make decisions, recognizes when it has finished and corrects its actions if needed, and halts and hands control back to the user if something fails. It also has access to tools to query outside systems and act on them, and it picks which one to use depending on the state of the task, within defined limits. The same guide rules out applications that plug in a language model without letting it control the flow: a simple chatbot, a single-turn answer or a sentiment classifier are not agents.⁠^{[1]OpenAI. A practical guide to building agents}

Anthropic draws the line in the same place, in different words. It calls anything that combines models and tools an agentic system, and inside that it separates two things: workflows, where the model and the tools follow paths defined beforehand in code, and agents, where the model dynamically directs its own process and its use of tools.⁠^{[2]Anthropic, engineering blog. Building effective agents} In practice, an agent works in a loop: it acts, looks at what the environment returns at each step (a tool's result, say) to see whether it is making progress, can stop to ask a person for help, and usually has a stopping condition, such as a maximum number of steps. Anthropic recommends starting with the simplest solution possible, because an agent trades latency and cost for better results on open-ended tasks.⁠^{[2]Anthropic, engineering blog. Building effective agents}

Tools are the piece that turns a model that talks into one that acts. The model's ability to request that a function run is tool calling, and the MCP (Model Context Protocol) standardizes how a server offers those tools to any application. Its specification defines them as model-controlled, discovered and invoked by the model according to context, and adds that there should always be a human in the loop with the ability to deny a call.⁠^{[3]Model Context Protocol. Tools (Model Context Protocol, specification 2026-07-28)}

When the task is buying, questions come up that a search does not have. Google sums them up when presenting its payments protocol for agents: proving that the user gave the agent authority for this specific purchase, that the agent's request reflects what the user wanted, and who is accountable if the transaction is fraudulent or wrong.⁠^{[4]Google Cloud, official blog. Announcing Agent Payments Protocol (AP2)} For buying inside ChatGPT, OpenAI describes its assistant as the user's AI agent, which passes information between the buyer and the merchant the way a personal shopper would.⁠^{[5]OpenAI. Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol}

In a store

A fictional example: a shopper tells an assistant, "I'm looking for a waterproof 30-liter backpack under $130 that arrives before Friday." An agent solves it in steps, and each step asks something different of the store:

If at step 1 the waterproofing is only in an image, the agent does not see it and the backpack is left out, even if it is the best answer. What the agent cannot read does not exist for it.

How to check

1. Ask for the list of tools. A real agent can show which tools it has and what each one does: read, write or pay.⁠^{[3]Model Context Protocol. Tools (Model Context Protocol, specification 2026-07-28)}
2. See who confirms. Before the agent does anything that changes data or moves money, there has to be a step where a person sees it and can reject it.⁠^{[3]Model Context Protocol. Tools (Model Context Protocol, specification 2026-07-28)}
3. Ask how it stops. An agent should have a stopping condition, such as a maximum number of steps, and hand control back when it gets stuck.⁠^{[2]Anthropic, engineering blog. Building effective agents}⁠^{[1]OpenAI. A practical guide to building agents}
4. Test it against your catalog. Give an assistant with search a shopping task in your category and note which of your store's data it used and which it could not find.

Related concept: agentic AI

Agentic AI is the field, and the agent is the concrete system. Anthropic uses "agentic systems" for the whole set, which includes both workflows with fixed paths and agents that decide their own path.⁠^{[2]Anthropic, engineering blog. Building effective agents} So not everything agentic is an agent: a process that summarizes reviews with a model and always follows the same steps is a workflow, even if it uses AI. For a store, the useful question is who decides the next step, the code or the model.

To keep going, agentic commerce applies all of this to buying, and the agentic commerce reading path lays out the protocols an agent uses to talk to a store.