AI Agent

Definition

An AI agent is a system given a goal, a set of tools, and the latitude to decide which tools to call and in what order to reach that goal.

The contrast that matters is with workflow automation, which follows a path defined in advance: if this, then that. A workflow handles the cases you anticipated. An agent handles the case you did not — the customer whose order shipped in two parcels, one scanned as delivered to a neighbour.

That flexibility is the entire value and the entire risk. It means the agent can also take an action you did not anticipate.

The three things that make it an agent

1. Tools. Functions it can call that change or read the world — look up an order, check stock, issue a refund. A model with no tools is a text generator, not an agent.
2. A loop. It observes the result of a tool call and decides what to do next, rather than executing a fixed sequence.
3. A stopping condition. It decides when the goal is met, or escalates.

A retrieval chatbot that answers product questions has none of these. It is useful, and it is not an agent.

Why the boundary matters more than the model

The most common production failure in agent systems is not a bad model output. It is an agent with more write access than anyone intended, discovered after it used it.

This is why in production the design order is: tool catalog first, authorization second, evaluation third, prompt last. Declare exactly which actions exist. Evaluate every write as a permission decision outside the model — a model can be argued out of a system prompt, but not out of an authorization policy. Then measure quality against a golden dataset with a pass bar before anything ships.

Autonomy is per action, not per agent

A single agent can reasonably be fully autonomous on reads, semi-autonomous on low-value writes, and gated on anything that moves money. Answering where an order is needs no approval. Issuing a refund does. Changing a price across a category definitely does.

Setting one blanket autonomy level for a whole agent is how teams end up either shipping something dangerous or shipping something so constrained it saves nobody any time.

When not to build one

• The decision is already deterministic. If a human follows the same rule every time with no judgement, that is a workflow. Workflows are cheaper, faster, and easier to audit.
• The data cannot support it. An agent reasoning over systems that disagree about identifiers produces confident, wrong answers. Fix the join keys first.
• Nobody owns the outcome. An agent without a named human owner accumulates edge cases until someone quietly switches it off.

Related terms

Tool catalog · Human in the loop · Agent evals · Model Context Protocol · Amazon Bedrock AgentCore · Agentic commerce