What Is An AI Agent? Workflows, Agents And When You Need One

“AI agent” is now applied to almost anything built on a language model, from chat assistants to scripted automations to systems that work for hours with nobody watching. When one label covers all of that, it becomes hard to judge what a supplier is offering or what your own team should build.

This article sets out the working definitions used by the companies that build the underlying models. It explains the loop that every agent runs, shows how agents differ from the fixed workflows that handle a great deal of everyday automation, and ends with a short decision guide for whether a task really calls for an agent.

Two Definitions That Agree

Anthropic’s engineering guidance separates two kinds of system built on large language models (LLMs).¹ In a workflow, a developer writes the path in advance: the model is called at set points and ordinary code decides what happens next. In an agent, the model itself decides which steps to take and which tools to use, and keeps control over how the task is carried out.

OpenAI’s guide for builders puts independence at the centre of its definition: an agent carries out work on someone’s behalf without being steered through each step. The guide names three parts that every agent has: a model that does the reasoning and makes decisions, tools the agent can call to fetch information or take action, and instructions that describe how it should behave and which limits it must respect.²

The wording differs but the core idea is the same. The test is who chooses the next step. If the code chooses, the system is a workflow, however capable each individual step may be. If the model chooses, within limits that people have set, the system is an agent.

The Agent Loop

Underneath, an agent repeats a simple cycle. It looks at its goal and at what it has learnt so far, chooses an action (usually a call to a tool such as a search, a database query or another service), and then reads the result. That result becomes new information for the next decision. The cycle carries on until the agent judges that the goal is met or something stops it.

This rhythm of reasoning, acting and observing was set out in the ReAct paper by Shunyu Yao and colleagues, first posted in 2022 and presented at ICLR 2023.³ The authors showed that letting a model alternate between written reasoning and actions helped it on question answering and interactive tasks. The loop described in this article follows that same basic shape.

The final step deserves more attention than it usually gets. OpenAI’s guide describes an agent run as a loop that continues until an exit condition is reached, such as a particular tool call or kind of output, an error, or a maximum number of turns.² An agent without a clear way to stop can repeat itself, run up costs, or keep acting long after it should have handed the task back to a person.

When A Simpler Design Is Enough

Both companies advise restraint. Anthropic recommends starting with the simplest approach that works, often a single well-designed model call, and adding agent behaviour only when it clearly improves the result. It also points out that agents accept higher cost and slower responses in exchange for better performance on hard tasks, so that exchange has to be worth it.¹

OpenAI suggests agents for three kinds of work that have resisted ordinary automation: decisions that need judgement and involve many exceptions, rule sets that have grown too tangled to maintain, and tasks that depend on reading unstructured material such as documents, emails or conversations.² Where none of these applies, a deterministic workflow is usually cheaper, faster and easier to check.

The flowchart below turns that advice into four questions. It is a starting point for discussion, and a team that answers it honestly will often find that a simpler design does the job.

What Gives An Agent Its Abilities

A model on its own produces output such as text; it cannot look up live records or change anything in another system. For tools that a developer defines, the model does not run anything itself: it returns a structured request naming the tool and its inputs, and the surrounding application decides whether to carry it out. Some providers also run a few built-in tools, such as web search, on their own infrastructure, in which case the controls are the ones that provider offers.⁴ Anthropic calls the basic unit of agent systems the augmented LLM: a model connected to retrieval so it can look things up, to tools so it can act, and to memory so it can keep track of what it has learnt.¹ The quality of those connections often decides how well an agent performs. Anthropic advises describing each tool carefully and testing how the model uses it, with the same care a team would give to a screen designed for human users.

Those connections are also what make agents consequential. A system that can send emails, change records or spend money needs permissions, monitoring and limits that a chat window never did. The security side of this is covered separately in Agentic AI Security. The next article in this series looks at tools, memory and the Model Context Protocol, the open standard many agents now use to reach outside data and services.