AI: Agents Terms

2026-10-16

About

Many people use the words "AI", "agent" and "AI agent" for very different things. A chat window gets the same name as a system that pays money. This post gives each word one meaning. It also gives one question that separates the words.

The examples come from the project in AI: Agents Reliable.

The question behind it

Somebody says: "We use an AI agent." What does that mean? It can be a chat window. It can be a fixed program with one LLM call. It can be a program that chooses its own steps and changes data. These three things have different risks. So they need different names.

The levels

Term	In one sentence	Example
AI	Generic term for software that learns from data.	spam filter, image recognition, ChatGPT
LLM (language model)	Text in, text out. Nothing else.	the model behind ChatGPT or Claude
Chat assistant	An LLM with a user interface. A person asks and reads.	ChatGPT in the browser
Workflow with an LLM	The steps are fixed. The LLM does some of the steps.	Clara: read (LLM), check (code), decide (code)
AI agent	A program in which the LLM chooses the next step and uses tools, in a loop.	Claude Code, Olive
Multi-agent system	Several agents or workflows share one task.	Olive, Penny, Clara, and Hugo through A2A

In the overview post, "loop" means agent and "graph" means workflow.

One question separates them

Who decides the next step: a person, the code or the model?

              Who decides the next step?
                          |
        +-----------------+-----------------+
        |                 |                 |
        v                 v                 v
   +---------+      +-----------+     +-----------+
   | person  |      |   code    |     |   model   |
   +---------+      +-----------+     +-----------+
        |                 |                 |
        v                 v                 v
  chat assistant    workflow with       AI agent
                      an LLM

        Figure 1: One question, three terms

A workflow is easier to test and to explain than an agent. So use a workflow where possible. Use an agent only where the steps are not known before.

Same technology, different use

Claude Code and Clara both use an LLM, tools and a loop or a graph. But they do different jobs.

	Coding agent	Business agent
Examples	Claude Code, OpenCode, Copilot CLI	Clara, Penny, Olive
Job	general: write code, edit files, run commands	one business task: check a claim, sort the mail
Tools	general: shell, files	business: read member data, pay, send a letter
Who uses it	a developer	a person in the business team

A developer uses a coding agent to build a business agent.

Parts that are not an agent

Some words name a part of an agent, not an agent:

• Harness: the program around the LLM: loop, tools, permissions, context. Agent = model + harness.
• Skill: a folder with a SKILL.md file. The agent reads it when the task needs it.
• MCP: the connection from an agent to a system.
• A2A: the connection from an agent to the agent of another company.

A tool like npx skills add only copies a skill folder into the folder that the coding agent reads. The agent gets new instructions to read. It does not get a new ability.

Four questions for each meeting

If somebody says "AI agent", ask these questions:

1. Who decides the next step?
2. Which tools can it use: only read, or also write?
3. Who approves a change before it happens?
4. How do you know that it works? What do you measure?