Why the Confusion Exists
Marketing departments do not have a strong incentive to use terms precisely. "AI agent" sounds more impressive than "AI assistant," so companies call things agents when they are not. "Agentic" appears in product documentation for tools that are simply chatbots with a few extra buttons. "Generative AI" gets applied to everything from a basic autocomplete to a fully autonomous system.
The result is that the same words mean different things in different contexts, and people trying to understand the space have no reliable vocabulary. That matters more than it might seem. These three categories have genuinely different capabilities, different risk profiles, and different use cases. Treating them as the same thing leads to bad decisions about what to build, what to buy, and what to worry about.
Here is what each term actually means.
Generative AI: Input, Output, Done
Generative AI is AI that produces content in response to a prompt. Text, images, code, audio. You ask, it generates, it stops. The model has no memory of the exchange after the session ends. It cannot take actions in the world. It does not check its own outputs against reality or try again if the result is wrong.
ChatGPT in a standard conversation is generative AI. Midjourney producing an image from a text description is generative AI. GitHub Copilot suggesting a line of code as you type is generative AI. Each of these tools is powerful and genuinely useful. None of them is doing anything except generating content in response to a prompt.
The key constraint: you control the loop. You give the input, you get the output, you decide what to do with it. The model is not taking any actions beyond producing text or an image. It cannot affect anything outside the conversation window.
Agentic AI: Input, Goal, Model Controls the Steps
Agentic AI adds one critical capability over generative: tool use. The model can call functions, observe the results, and decide what to do next based on what it sees. It is not just generating content. It is taking a sequence of actions to reach a goal.
A clear example is Claude Code working on a software project. You give it a goal: add a feature to this codebase and make sure the tests pass. The model reads files, writes code, runs the tests, reads the output, fixes errors, runs the tests again. It is making decisions at each step based on what it observes. You did not specify each action. You specified the goal.
The loop has moved inside the model. In generative AI, you decide what to do with each output and whether to ask the next question. In agentic AI, the model decides what the next action is based on what the previous action returned. You set the goal. The model figures out the steps.
This is a real change in capability. It is also a real change in risk. The model can take actions in the world: write files, call APIs, send requests to external services. If it makes a wrong decision, the consequences are not contained to a chat window. They happen in whatever systems the model has access to.
Another useful example is a research agent that you ask to gather information about a topic, search the web, read source documents, and compile a structured summary. You did not direct each search query or specify each source. The model did that work based on a goal. The session ends when the task is complete or when you stop it. Nothing persists after the session closes.
AI Agents: Goal, Constraints, System Runs Across Time
AI agents add two more things over agentic AI: persistence and autonomy across sessions.
An agentic AI runs during a session and stops when the session ends. An AI agent keeps running. It has memory that persists between sessions. It can maintain context files that it updates over time. It takes scheduled actions without human initiation. It manages ongoing goals across days or weeks or months.
A Hermes agent monitoring your inbox is an AI agent. It runs on a schedule. It remembers what it has seen before. It takes actions (flagging emails, drafting responses, updating a context file) without you initiating each one. You defined the goal and the constraints. The system runs the loop across time without your involvement between cycles.
The practical difference from agentic AI is that you are not watching. With agentic AI, you give a goal and observe the result before deciding what to do next. With AI agents, the system completes cycles on its own. Your review might be weekly, not moment-to-moment. The actions the agent takes between your reviews are real and may be difficult to reverse.
The Risk Ladder
These three categories are not just different in capability. They have different risk profiles, and understanding the difference matters for deciding how to deploy each one.
Generative AI can produce bad outputs. Incorrect information, harmful content, poorly written code. The damage is contained to whoever reads the output and decides to act on it. A human is always in the loop before anything happens in the world.
Agentic AI can take bad actions within a session. If a coding agent writes and runs code that deletes files, the files are gone. If a research agent calls an API too many times, real costs accrue. The model is taking real actions, and some of those actions may be difficult or impossible to undo. The risk is bounded to a single session, but it is not hypothetical.
Autonomous agents can take bad actions over extended periods and be harder to reverse. An agent running for weeks or months with write access to external systems can accumulate errors, take actions based on stale context, or drift from its original goal in ways that compound over time. The damage can be larger and more spread out than a single bad session, and the trail of what the agent did may not be easy to reconstruct.
This does not mean agents should not be used. It means the safeguards appropriate for each category are different. A generative AI chatbot needs content moderation. An agentic coding tool needs sandboxing and the ability to roll back file changes. An autonomous agent needs clear boundaries, a review mechanism, and logging of every action it takes.
The Spectrum, Not the Boxes
These three categories are useful conceptual anchors, but the reality is a continuum. Most products marketed as "agents" today sit closer to the agentic end than the truly autonomous end. They have some persistence, some tool use, and some ability to take scheduled actions, but they still require significant human oversight to function well.
Fully autonomous agents, systems that genuinely pursue goals across time without human intervention between steps, exist but are rare in production deployments. The capability exists. The infrastructure for running them reliably, including the monitoring, logging, and intervention tools that make them safe to run autonomously, is still maturing.
When a company tells you their product is an "agent," the useful questions are: does it maintain memory across sessions, does it take actions without human initiation between those actions, and what happens if it makes a mistake? Those questions tell you where on the spectrum it actually sits.
The framing that helps most: think about where human oversight happens. In generative AI, oversight is continuous because you are making the decisions. In agentic AI, oversight happens at the start (you define the goal) and at the end (you review the result). In autonomous agents, oversight is periodic and after the fact. The further you move toward autonomous, the more the oversight model has to change to match the risk.
The vocabulary matters because the decisions are different.
A generative AI wrong answer needs a correction.
An agent wrong action needs a recovery plan.
Knowing which one you are dealing with is where good deployment decisions start.