What "agent" means when a company says it
The word covers four different things, and companies rarely say which one they mean. Here is how to tell them apart in about ten seconds, and what actually changed to make the word worth using.
Ask two questions: does it act without asking permission each time, and does it loop until done? Most things marketed as agents answer no to both and are really tool-calling chatbots.
"Agent" went from a research word to a marketing word in about eighteen months. It now appears in press releases from companies whose product is a chatbot with a longer prompt. The word has not become meaningless, but it has become four words wearing one coat, and nobody tells you which one they are selling.
The four things people call agents
One: a chatbot with a personality. A model with instructions telling it to behave a certain way. No actions, no tools. This is a prompt. Calling it an agent is marketing.
Two: a model that can call tools. The model can look something up, run a search, hit an API. This is genuinely useful and it is what most products mean. The model is still answering one question at a time, and a human is still in the loop for every turn.
Three: a loop. The model plans, takes an action, looks at what happened, and decides what to do next, repeating until the goal is met or it gives up. This is the first one that deserves the word. The defining feature is that nobody knows in advance how many steps it will take, including the system.
Four: a standing autonomous system. A loop that runs without being asked, has its own permissions and budget, and keeps going when you close the tab. Rare, and the interesting problems are almost all about safety and money rather than intelligence.
The ten-second test
When someone says agent, ask two questions:
- Does it take actions in the world without a human approving each one?
- Does it loop, deciding its own next step, until it is done?
Two yeses is an agent in the useful sense. One yes is tool-calling. Two noes is a chatbot. You can usually work it out from a product demo faster than from the documentation.
What actually changed
Two real things, underneath the noise.
Models got reliable enough at using tools that a loop stops being a toy. An agent that is 90% reliable per step is worse than useless over twenty steps, because 0.9 to the power of 20 is about 12%. Compounding is brutal, and it is why agents felt like a demo for years. Every point of per-step reliability moves the whole curve.
Tool access got a standard. Anthropic published the Model Context Protocol in November 2024 and it has been adopted well beyond them. Before a standard, every tool needed bespoke wiring to every model. This is the boring infrastructural thing that actually unlocks the category, in the same way USB mattered more than any individual peripheral.
Where it genuinely does not work yet
This is the part the press releases leave out, and the part that makes you sound like you know what you are talking about.
Long-horizon tasks remain hard. Agents lose the thread, repeat themselves, declare victory early, and fail in ways that are difficult to notice because the output is fluent. Fluent and wrong is a much worse failure mode than obviously broken, and it is specific to this technology.
Benchmark scores on realistic multi-step work are far below what the marketing implies. I am deliberately not quoting a single headline number here, because agent benchmarks move fast, differ wildly in what they measure, and are easy to quote misleadingly. If you want to look, SWE-bench is the most watched for software tasks and publishes its own leaderboard. Read the task definition before you read the score, because the definition is doing most of the work.
The honest summary
The word is worth keeping. A system that decides its own next step is genuinely a different kind of thing from a system that answers a question, and we needed a word for it.
But when a company says agent, they usually mean number two. That is not a scandal and tool-calling is useful. It is just not autonomy, and the gap between those two is where most of the disappointment in this category comes from.
Found something wrong here? That is worth more to me than a compliment. Tell me and it gets corrected on the page, with the date.