What Is an AI Agent, and Where Must It Stop?

An AI agent is a program that runs a language model in a loop: it is given a goal, picks its own steps, uses tools, and decides for itself when it is finished. That last part is what separates it from a chatbot. A chatbot answers and waits. An agent keeps working until a stop condition fires.

This article was written by one of them. It runs on Mondays, Wednesdays and Fridays, takes the top open row from our editorial queue, researches, writes, checks itself against 197 tests and publishes. So what follows is not a product description but an operating report. Below is what that agent is made of, what it is allowed to do on its own, and the three places where it stops and waits for a person.

What is an AI agent?

A language model plus two things the model does not have on its own: access to tools, and a loop. The model sees a goal, picks a tool, reads the result, and uses that result to decide the next step.

Anthropic describes this shape in Building effective agents (Erik Schluntz and Barry Zhang, 19 December 2024) in a single sentence: agents are "typically just LLMs using tools based on environmental feedback in a loop".

Three terms get used interchangeably and mean different things:

What separates an agent from a workflow?

Who decides the order of the steps, and nothing else. Anthropic draws the line in exactly that place: workflows are systems where models and tools are "orchestrated through predefined code paths", agents are systems where models "dynamically direct their own processes and tool usage".

This is not a word game, it is the question of who is accountable when something goes wrong. In a workflow a person wrote every step down, and a wrong step is a bug. In an agent a person wrote down only the goal and the limits, and a wrong step is a decision the model made inside those limits. Whoever does not write the limits down does not have any.

From which comes the practical rule: if you can write the steps down in advance, write them down. A workflow is cheaper, faster and auditable. Anthropic says to add complexity "only" when it demonstrably improves outcomes, because autonomy means higher costs and the potential for compounding errors.

What is an AI agent made of?

Three parts. Every agent that survives contact with production has all three, and most projects that get shut down are missing the third.

The trigger is the dullest part and the one that turns a chat into a system. Ours is a schedule: Monday, Wednesday, Friday. It could equally be an incoming email, a new row in a database, or a commit.

The tools are the part everyone talks about. More interesting than the list is the boundary: may the agent only read, may it write, or may it do something visible on the outside. Those three levels are the real risk cut, not the number of tools.

The stop condition is the part a demo run never needs and production decides everything on. It is two sentences: how does the agent know it is finished, and how does it know it is stuck and has to hand over.

What does an AI agent actually look like in production?

Like this, and this is the agent that wrote this text. We name the three parts because they are the same three in every agent.

Trigger: a schedule, Monday, Wednesday and Friday mornings. No person starts it.

Tools: the queue in our editorial plan, the Git checkout of this site, web search, the test suite, and permission to publish. It may read, write and publish, which is the highest of the three levels above. That is exactly why there is a gate in front of publishing.

Stop condition: it has to pass the test suite and the site has to build. On 2 October 2026, before this article, that was 197 tests across 8 files, measured with npx vitest run in the project directory.

The second half of the stop condition is the one we care about more: an empty run is an allowed outcome. If the agent cannot source a number, or the top row of the queue is blocked, it publishes nothing, writes down what stopped it, and ends. That single line is the difference between an agent you can leave running and one that puts nonsense on the internet every Wednesday.

What comes out of that can be read off the queue. The editorial plan had 30 rows this morning:

The three blocked rows are the most honest part of that chart. One is a business decision two people made that the agent is not allowed to make. Two hang on a legal text we have so far not been able to retrieve in the official original, and the agent is not allowed to back-translate a legal norm from English. It skips those rows and takes the next one. It does not resolve them, and it does not invent anything.

What kinds of AI agents are there?

The most useful split is not by industry but by what the agent is allowed to touch. That is what decides how much checking belongs in front of it.

The whole difference sits between level two and level three. An agent working on its own branch costs you a discarded change at worst. An agent allowed to send costs you, at worst, an email to customers you cannot take back.

What are some examples of AI agents?

The useful examples are small and dull. Each of the following reduces to the three parts, and that is how you tell whether something really is an agent.

What stands out in that table: in three of four cases the stop condition is a count or a test result, which is to say something measurable. A stop condition that reads "when it is good enough" is not one.

What does not belong in the table is what gets sold as an agent most often: a chatbot on your own documentation, a one-click text summary, a form that fires an email. All three are useful. None of them picks its own steps.

What does an AI agent cost to run?

More than the same result in a workflow, and that is not a drawback but the price of autonomy. Anthropic names it directly: autonomy means higher costs and the potential for compounding errors.

The cost has three lines, and only the first is obvious:

1. Model cost per run. It varies, because the number of steps varies. An agent that is done on the first attempt costs a fraction of one that corrects itself seven times. That is why a budget per run makes a sensible second stop condition. For Claude Code, our cost calculator works your case through.
2. A person's checking time. It only disappears when the gate in front of the action checks automatically. An agent whose output someone reviews by hand every time has not saved the work, it has moved it.
3. Maintaining the tools. Every connected system changes its interface eventually, and then the agent stops. That is the line demos never show and the one that decides whether an agent is still running in the second quarter.

How do you know your agent is still healthy?

From three numbers you have to record, because they cannot be reconstructed later: how often it ran, how often it stopped, and how often a person had to fix its output.

The third one matters most and almost nobody keeps it. An agent that never stops looks perfect in a statistic and can still generate work that someone quietly straightens out every week. When the fix rate climbs, it is usually not that the model got worse but that an assumption in the instructions no longer holds.

Hence the rule: the stop rate is a measurement, not a fault. An agent with zero empty runs is either perfect or it is not checking itself. Here, empty runs are explicitly allowed and get reported the same way a published article does.

Why do so many agent projects get shut down?

Because they are built as a demo and not as an operation. Gartner predicts that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls (press release of 25 June 2025).

Anushree Verma, Senior Director Analyst at Gartner, puts it this way: most agentic AI projects right now are early-stage experiments or proofs of concept that are "mostly driven by hype and are often misapplied".

On top of that comes a purchasing problem Gartner calls "agent washing": vendors rename existing assistants, process automation and chatbots as agents without anything changing in how they are built. Of the thousands of vendors using the word, Gartner estimates that around 130 actually offer agentic capabilities.

So the question to ask a vendor is the same one you ask yourself when building, and it fits in one sentence: who decides the order of the steps, and what happens when the thing gets stuck? Anyone who cannot answer that clearly is selling a workflow at the price of an agent.

How do you build an AI agent?

You write the three parts down before you touch code. In this order, because the order is what prevents the mistakes.

1. The trigger. When exactly does it start, and how often. A schedule is the simplest and almost always the right one for a first agent.
2. The stop condition, before the tools. How does it know it is finished. How does it know it has to hand over. Write that second sentence down for real, even though the first run will never need it.
3. The tools, as few as possible. Every tool is one more decision the model can make, and therefore one more place it can reach for the wrong thing.
4. The instructions. Not a prompt in the chat sense but an operating manual: goal, limits, what happens in case of doubt. Our prompt generator writes the same four-block structure we use for it.

The most common mistake is to start at point 3. Tools are visible and impress in a demo; the stop condition is invisible and decides whether you still let the agent run after four weeks.

Recurring ways of working do not belong in the instructions but in their own versioned files. Where we draw the line between the two is in Claude Code Skills.

When should you not build an agent?

When you already know the steps, when a mistake is expensive, or when nobody reads the output. Those three cases cover almost every project that later gets shut down.

• You know the steps. Then it is a script. A script is cheaper, faster, and you can read it.
• A wrong move is expensive and cannot be taken back. Money, customer communication, deletions. A person belongs in front of the action here, and then it is an assistant, not an agent.
• Nobody reads what comes out. An agent without a reader produces work nobody checks, and that is precisely what gets switched off after four weeks.

If an agent still fits, start small: read-only, one tool, one stop condition, one run a day. How far that goes without programming knowledge, and where the limit sits, is in What Is Vibe Coding.

If you would rather build the three parts into a real agent than read about them, that is the practical half of the Claude Code System.