Introduction

The term 'agentic' come from the Latin word 'agere' (meaning "to do," or "to act). It includes the noun agent (agens, "one who acts").

According to many Computer Scientists, tool use is the defining feature of an agentic system. Researchers during the ‘80s and ‘90s used the word “agent” to describe an AI system that could “act” (using tools). In 1995, Russell and Norvig wrote in their massively popular Artificial Intelligence: A Modern Approach, that Agents are a class of AI systems that can take actions with varying degrees of autonomy, and use tools, in order to pursue goals.

Eventhough the conceptual framework was already largely there, what an 'agent' was capable of remained limited by the capabilities of the available AI systems in their time. It has taken decades for machine learning to develop to the point where agents could have finally become useful tools.

But when do we call something an 'Agent'? The boundary between a 'deterministic workflow', an 'LLM-augmented workflow', an 'LLM-Agent' and an 'Autonomous LLM-Agent' often gets murky. I will try to split up how I am using these terms throughout this blog, and which different industry definitions they refer to.

Traditional Workflows

Using tools can be very general and abstract. Often this is not meant in the sense of a robot physically using a tool, but an AI system using some other computer system (like a calculator, or any other kind of software). But a system that follows some deterministic rules to call a series of tools is the essence of what any piece of software is doing.

Tool-Augmented LLM (or LLM-Augmented Workflow)

When we include LLM as part of a workflow, we can create workflows that allow us to process much more general inputs. If we think about these structures as essentially being workflows, we can more accurately refer to them as LLM-Augmented workflows.

Alternatively we might think of them as Tool-Augmented LLMs. This is Anthropic's "augmented LLM" definition (LLM + retrieval + meory + tools). But since LLM + Tools also meets the definitions used throughout the ‘80s and ‘90s, where an “agent” is described as an AI system that could “do” something. An LLM-Augmented Workflow if often equated with an 'AI Agent'.

The Agentic Loop

The spectrum from 'workflow' to 'agent' is the degree of autonomy that we delegate to a system to pursue a goal.

As far back as 1948, Norbert Wiener’s introduced the idea of a feedback loop in his book Cybernetics: Or Control and Communication in the Animal and the Machine. Such a feedback loop enables an entity to sense their environment, process information and adjust behavior accordingly, in order to persue a goal.

This theoretical framework is still very close to how we think about AI agents and their architecture today. By including a loop in an 'LLM-Augmented Workflow', we allow such a system to continue running until it is satisfied that it has achieved its goal and exits the loop.

This is the definition that most of the AI-engineering community converged on in 2025: "An LLM agent runs tools in a loop to achieve a goal." Essentially this is an LLM-Augmented Workflow run in a loop. And the distinction between these two, and which one is called an agent, is often where most of the confusion comes from.

Some groups insist that Agentic tool use is the defining feature of an agent. While other groups focus on the loop-based persuing of a goal as the defining feature.

Autonomous Agents

We can generalise things even further, and instead of having a workflow (that can choose between a fixed series of tools), allow such a system to create its own tools that it can use again within the Agentic Loop that its running.

This setup (an agent that can create its own tools) is the basic pattern behind 'Autonomous' AI Agents like Claude Code (at least when you run it with full permissions).