The short version
- What an agent is. A software program that uses an AI model as one part inside it. The model reasons; the software around it holds the goal, picks the next step, uses tools, and repeats until the work is done.
- The AI is one part. Swap the model out and the agent still runs. The rest -- instructions, the loop, tools, memory, guardrails -- is ordinary software.
- How to tell one apart. Once it has a goal, an agent chooses and runs its own steps. A chatbot, a plugin, or a workflow waits for a person or follows a fixed path.
What is an AI agent?
Everyone has heard of AI agents. Many people have even built them, and nearly every software company claims to offer them, yet few can clearly explain what they are. For starters, an AI agent is not simply AI. It is a software system that uses one or more AI models as part of how it works.
A model on its own receives input, generates an answer, and stops. An agent surrounds the model with instructions, memory, tools, permissions, and a program that manages the work. That system can hold a goal, choose the next step, use tools such as web browsers, calculators, files, and connected applications, evaluate what happened, and continue until the work is finished or human input is required. The model supplies much of the reasoning, while the surrounding system turns that reasoning into action.
A program, not a model
The word "AI" in "AI agent" points at the model. The word "agent" points at the program built around it. That program decides what to do next without a person approving every step, and that is the whole difference. A chatbot waits for the next prompt. An agent does not. It looks at where it is, chooses an action, takes it, checks the result, and goes again.
This is why the model and the agent are not the same thing. Swap the model out and the agent still works. Strip the agent down to the model alone and it stops being an agent. Anthropic draws the same line: a workflow follows predefined code paths, while an agent "dynamically directs its own processes and tool usage," typically behaving as "LLMs using tools based on environmental feedback in a loop" [1].
The parts of an agent
Six parts make up a typical agent. Only one of them is AI.
- Instructions (its programming). The goals and rules written in plain language. This is the agent's job description, and it shapes every decision the agent makes.
- Orchestrator and loop. The code that runs the loop: it gives the model the current state, takes back the model's proposed action, checks and runs it, and repeats until the goal is met or the budget runs out. This loop is what separates an agent from a single answer.
- LLM (generative AI). The AI part. It reads the current situation and proposes the next action. It may be one model or several working together, and it is still one component inside the program.
- Tools. The connections that let the agent act in the world: APIs, web browsing, code, and files. Tools are what turn a proposed action into an effect.
- Memory. Working state for the task in front of it, plus longer-term notes it carries forward. Memory is what lets an agent build on what it already did instead of starting over.
- Guardrails. Permissions, logging, and monitoring that bound what the agent is allowed to do and record what it did. Guardrails run across every other part.
Three levels of agent
Agents are not one thing. They arrive in three levels, each built on the one before it: a basic agent that answers, a tool-using agent that acts, and a proactive agent that keeps working over time. The same six parts appear at every level. What changes is how many of them are in play.
What an agent is not
Many things marketed as agents are not agents. They are capabilities and connections that act only when called.
- A model or prompt generates and reasons over text. It has no memory, no tools, and no actions. It does nothing until it is asked.
- A skill is a repeatable procedure loaded into a host when it is relevant. It applies a method. It does not decide to act [2].
- A plugin or custom GPT bundles skills and connections into a chat surface. It brings capabilities into the conversation, but it does not pursue a goal on its own [3][4].
- A connector or MCP server is a standard link to a live system such as email, a calendar, or a database. It moves information. It does not choose what to do with it [5].
None of these decides and acts by itself. Only two things in the stack do: a single agent, and a multi-agent system where agents coordinate and split the work. Every claim about AI agents is about those two, and so is every risk.
The building blocks, compared
The same terms appear across ChatGPT, Claude, and the rest of the stack, and they do not mean the same thing. The table below sorts them, from the pieces that act only when called up to the two that choose their own steps. Read it top to bottom.
| Type | What it is | Who builds it | What it can do | Chooses its own steps? | Example | Watch-out |
|---|---|---|---|---|---|---|
| Chatbot | A conversational surface for a model. Text in, text out, turn by turn. | Vendor | Answer and converse | No | ChatGPT, Claude, Gemini in plain chat mode | Feels agentic but waits for every prompt |
| Model / prompt | A raw model call. Text in, text out. | Model provider | Reason over text | No | Asking a model one question | No memory, no tools |
| RAG pipeline | Retrieval plus a model: find the relevant passages, then answer from them. | You | Answer from your own data | No | A docs assistant over a wiki or a knowledge base | A fixed pipeline, not an agent -- sometimes just a model call with no agent at all. When an agent does drive it, the agent is the part that decides. |
| Skill | A capability: a method plus instructions and scripts, loaded when relevant [2] | You (portable) | Apply a repeatable procedure | No -- only when invoked | A security-review method | Only as good as its instructions |
| Plugin / GPT | A packaged surface you add to a chat; bundles skills, tools, and connectors [3][4] | Vendor / community | Bring capabilities into a chat | No | A custom GPT; a ChatGPT plugin | Naming churns (GPTs to Plugins in 2026) |
| Connector / MCP server | A standard connection to a live external system [5] | Vendor / you | Read and write external state | No | A Gmail connector; a GitHub MCP server | Real-world access -- scope it tightly |
| Workflow / automation | A fixed path of steps a developer writes in advance. | You | Run a defined sequence | No -- follows its code path | A nightly sync; an approval flow | Brittle when conditions change |
| Copilot | An assistant embedded inside a tool, working alongside a person. | Vendor | Assist inside a fixed surface | No -- the person stays in the loop | A code assistant; a writing assistant | Branding spans chat to agent -- check the mode |
| Agent | A model plus instructions, tools, memory, a loop, and a goal [1] | You / vendor | Choose and run their own steps | Yes | A research-and-write assistant | The leverage point -- and the risk |
| Multi-agent system | Agents that spawn or coordinate with other agents. | You | Split and coordinate work | Yes | Sub-agent teams | Shared state becomes a channel |
The first eight rows are capabilities, connections, and fixed paths. They act only when called. Only the last two rows choose their own steps. Every headline about "AI agents" is about those two rows, and only those rows carry the risks that come with autonomy.
What differentiates an agent
When a product, a feature, or a demo arrives with the word "agent" on the label, ask one question: once it has a goal, does it choose and run its own steps, or follow steps set by a person or a fixed path? If it chooses its own steps, it is an agent. If a person sets the steps, or the path is fixed, it is not.
That single question does most of the work. It shows where the leverage is, which is the parts that act, and it shows where the controls belong, which is the same parts. An agent that chooses and runs its own steps is where guardrails, memory you can inspect, and logs you can read matter most.
Is ChatGPT an AI agent?
ChatGPT is a software product that can host several kinds of AI interaction. Depending on the mode and task, it may operate as:
- A chatbot that generates a response
- A tool-using assistant that searches or analyzes files
- An agent that performs a multistep task through repeated model and tool calls
- A platform that hosts configured workspace agents
A simple chat response is not necessarily agentic. If you ask, "Rewrite this paragraph," ChatGPT may call the model once and return its output. If you ask it to research a topic, inspect files, execute code, evaluate results, and continue until it reaches an answer, ChatGPT is operating as an agent. OpenAI describes this behavior as a sequence of model requests and tool calls coordinated by an agentic harness.
So the most precise statement is: ChatGPT is an AI application with agentic capabilities. In some modes and tasks, it functions as an AI agent.
Agents do not have to come from a chatbot
An AI agent does not have to come from ChatGPT, Claude, or another general-purpose assistant. It can be built into specialized third-party software.
Buildr is a preconstruction platform for general contractors, and Kit is its AI agent. Buildr describes Kit as an agent that can read emails and project documents, evaluate information, update CRM records, draft proposals, flag conflicts, and use the result of one step to decide what to do next [6]. Based on that published description, Kit reasonably qualifies as an AI agent. Buildr is the platform that contains and operates it.
The same distinction applies to products that let customers create their own agents:
- The agent builder is the platform.
- The configured worker you create is the agent.
- The underlying LLM is the AI model used by that agent.
- Connected applications, APIs, calculators, browsers, and databases are its tools.
Not everything marketed as an "agent" behaves like one. The test is whether it receives a goal, chooses among possible steps or tools, observes the results, and adjusts what it does next. If it does, it is reasonably an agent, even if its job is narrow. If it simply follows a fixed sequence, such as "when an email arrives, summarize it and enter the result in this field," it is better described as an AI-powered automation or an agentic workflow.
There is therefore a spectrum: LLM feature → AI automation → constrained agent → general-purpose agent → multi-agent platform.
A narrowly configured Buildr worker can be a genuine agent without being as broadly capable as Muse, ChatGPT, Claude, or an OpenClaw platform [7]. "Agent" describes how the system operates, not how large, independent, or intelligent it is.
Why the distinction matters
The difference is not academic. Everything a person drives step by step, or a fixed path a developer wrote, is safe by default, because a person starts it or the path bounds it. An agent, once it has a goal, chooses and runs its own steps inside its loop, so the same design that makes it useful also makes it the part that needs bounds. Define the goal tightly, grant only the permissions it needs, and watch what it does. That is the whole design problem, and it is the same whether the agent runs one task or a thousand.
At FutureInSites, this is where every agent engagement starts: agree on what an agent is, then bound what it can do. Clarity here carries through everything you build on top of it.
Agents choose and run their own steps toward a goal, which raises the next question: whether that can make them dangerous, and under what conditions. That is the subject of the companion post, Are AI Agents Dangerous?
References
- Anthropic, "Building effective agents" (Dec 19, 2024). anthropic.com
- Anthropic, "Skills explained: How Skills compares to prompts, Projects, MCP, and subagents." claude.com
- OpenAI, "Introducing AgentKit." openai.com
- OpenAI, "Custom GPT retirement and migration FAQ," OpenAI Help Center (Custom GPTs scheduled to retire Dec 11, 2026). help.openai.com
- Anthropic, "Introducing the Model Context Protocol" (Nov 25, 2024). anthropic.com
- Buildr, "Kit, Buildr's AI Agent" (product page). buildr.com
- Meta, "Introducing Muse: The World's First Personal AI Agent Built for Everyone" (Sep 8, 2026). about.fb.com