In plain terms
You tell ordinary software how to do something. You tell an agent what you want. It works out the how, does it, checks whether it worked, and tries again if it did not.
TL;DR
- There is no standard definition. Anthropic's engineering guidance states plainly that the term is used in several ways, and groups the whole family under agentic systems rather than picking a winner.
- Anthropic's line: workflows orchestrate models and tools through predefined code paths; agents are systems where models dynamically direct their own processes and tool usage.
- OpenAI's line is broader: systems that independently accomplish tasks on your behalf.
- They agree on one thing, which is the useful part: autonomy over the sequence is what separates an agent from everything else.
- Anthropic's own advice is to start simple and add agency only when a simpler approach falls short. That is the company that sells agent-building tools saying you probably need less than you think.
If you are already weighing an agent against other options for a specific process, the sorting question is covered in AI agent vs chatbot.
What is an AI agent?
An AI agent is software given a goal rather than a fixed set of instructions, which decides for itself which steps to take, uses tools to take them, and continues until the goal is met or it stops to ask a person.
Three plain-English terms that make the rest readable:
- A large language model, or LLM, is the kind of artificial intelligence that reads and writes text. It is the part that reasons. On its own it can only produce words.
- Tools are the other software an agent is allowed to use: a database, an email system, a calendar, a search engine. Tools are what let it act rather than only describe.
- Autonomy is how much of the sequence the software decides for itself. This is the variable that every competing definition is really arguing about.
A note on wording, because it matters below. Anthropic uses agentic systems as the umbrella term covering everything in this family, then distinguishes agents from workflows inside it. That is the more precise vocabulary and almost nobody uses it correctly.
In plain terms, what does one actually do?
Take a concrete task: researching a company before a sales call.
Ordinary software would need you to specify every step. Open this database, search this field, copy this result, paste it there.
An AI agent gets the instruction "research this company and write me a one-page brief". It decides to search the web, reads what it finds, notices the company announced funding last month, decides that is relevant, looks for more detail on it, checks your CRM for previous contact, and writes the brief. Nobody told it to look for the funding announcement. It decided that was worth doing.
Run the same task twice and it may take different routes. That flexibility is the whole value, and it is also the whole risk, because you cannot fully predict what it will do.
Why do definitions of AI agent disagree?
Because the term arrived in marketing before it settled in engineering, and the major labs have drawn the line in different places.
Anthropic is unusually direct about this. Its engineering guidance, Building effective agents by Erik Schluntz and Barry Zhang, published December 2024, opens by acknowledging that the term is used in several ways: some people mean fully autonomous systems operating independently over long periods, others mean more prescriptive implementations following predefined workflows. Anthropic groups all of it under agentic systems, then draws this architectural distinction:
| Definition | |
|---|---|
| Workflows | Systems where language models and tools are orchestrated through predefined code paths |
| Agents | Systems where language models dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks |
Anthropic describes agents further: they begin with a command from, or a discussion with, a human user. Once the task is clear, they plan and operate independently, potentially returning to the person for information or judgement.
OpenAI defines an agent more broadly, as systems that independently accomplish tasks on your behalf.
Where they agree is the part worth keeping: autonomy over the sequence. Both definitions turn on whether the software decides its own steps.
Why this matters commercially. A vendor calling a fixed, rule-based automation an "AI agent" is using a definition loose enough to permit it. They are not necessarily lying. The word does not have a fixed meaning, which is exactly why it appears in so much marketing. You need a test rather than a definition, and section five is the test.
What is an AI agent not?
Five things it gets confused with.
| What that actually is | |
|---|---|
| A chatbot | Software you talk to that returns words. It answers; it does not act on other systems |
| A workflow automation | Software running steps you defined, in the order you defined them, identically every time. Under Anthropic's terms this is a workflow, not an agent |
| A large language model | The reasoning component. An LLM on its own has no tools and cannot do anything except produce text |
| Robotic process automation | Older software that clicks through screens on your behalf, following a recorded path. No judgement involved |
| Artificial general intelligence | A hypothetical system with broad human-level capability. Current agents are narrow, task-specific and supervised |
The most common confusion is the second one. A great deal of what is marketed as an AI agent is a workflow with a language model called at one step. That can be an excellent system and frequently the right one. It is just not what Anthropic's definition describes, and the difference shows up in your bill and your predictability.
The one question that settles it
Who controls the sequence of steps?
Ask a vendor to show you two runs of the same task. If both runs took the identical route, the software followed a path somebody wrote. If they took different routes because the software decided differently based on what it found, it is directing itself.
| You control the sequence | The software controls the sequence | |
|---|---|---|
| Anthropic calls it | Workflow | Agent |
| Same input, same steps? | Yes, always | Not necessarily |
| Cost per run | Fixed and predictable | Variable |
| Easy to audit afterwards | Yes, the steps are written down | Harder, the route changes |
| Best for | Anything with a known sequence | Variable inputs where the route depends on what is found |
This test works regardless of whose definition the vendor is using, which is why it is more useful than the definition itself. It is also the same question you should ask before deciding what to build, which our AI agent vs chatbot guide applies across three technologies.
What are the parts of an AI agent?
Four components. Every agent has them, whatever the vendor calls its architecture.
1. The model. The reasoning engine, usually a large language model. It decides what to do next.
2. Tools. The other software it may use. A database, an email system, a search engine, your CRM. The connections matter more than people expect, which is why a standard for them emerged: the Model Context Protocol, announced by Anthropic in November 2024 as an open way for models to connect to external data and tools through a single interface.
3. Memory. What it retains. Within a single task, so it does not repeat itself. Sometimes across tasks, so it learns your preferences.
4. The loop. The part that makes it an agent rather than a single question. It acts, observes the result, decides whether the goal is met, and either continues or stops. Anthropic's guidance describes this as the agent planning and operating independently until a stopping condition, potentially returning to the human for judgement.
Guardrails sit around all four. These are the limits you set: which systems it may touch, how much it may spend, and which actions need a person to approve them. On anything consequential they are not optional.
What can AI agents actually do today?
Narrow, supervised tasks with messy inputs and checkable outputs. The pattern is consistent.
| Task | Why it fits |
|---|---|
| Reading an incoming support ticket, classifying it and drafting a reply | Input varies enormously, output is reviewed before sending |
| Researching a company and producing a brief | The route depends on what it finds; a person reads the result |
| Extracting fields from invoices or delivery notes | Documents arrive in dozens of formats, extraction is checkable |
| Summarising interviews and clustering themes | Unstructured input, no single correct answer |
| Writing and testing code against a defined issue | The route varies, and tests provide an objective stopping condition |
In every case the input is messy, a person uses the output, and being wrong occasionally is survivable. That is the shape that works today.
What they cannot reliably do: anything requiring guaranteed identical behaviour, anything where a wrong answer moves money without review, and anything you cannot describe well enough to check. Those are workflow territory.
When should you not use one?
When the steps are the same every time. If you can write the sequence down, write it down. A workflow is cheaper, faster and behaves identically on every run.
When you need predictable costs. Agents are charged per token, roughly a chunk of a word, and the amount varies because the agent decides how much thinking to do. A workflow's cost is fixed per run.
When you need to explain exactly what happened. Auditing a fixed sequence is reading a list. Auditing an agent means reconstructing a route it chose.
When nobody will review the output. Agents can be confidently wrong. Without a person checking the cases that matter, that is a liability rather than an efficiency.
Worth noting who says this. Anthropic's own guidance advises starting with simple prompts and adding agentic systems only when simpler approaches fall short. That is the company selling agent-building tools telling you to use fewer of them. It is the most credible advice on this page and it is not ours.
If the honest answer is that your process has a fixed sequence, business process automation is the relevant discipline, and it is considerably cheaper.
What do people get wrong about AI agents?
Assuming the word means one thing. It does not, and the labs say so. Ask what the vendor means before evaluating the claim.
Calling a workflow an agent. Extremely common, and often harmless, but it changes the cost model and the predictability. Ask the two-runs question.
Expecting deterministic behaviour. Same input, same output is a property of workflows. An agent that always took the identical route would not need to be an agent.
Skipping guardrails. Spending caps, approval points and access limits are what let you keep an agent after its first surprise.
Confusing agents with general intelligence. Current agents are narrow, task-specific and supervised. The gap between "handles support triage well" and "thinks" is not a small one.
Believing more autonomy is better. More autonomy means more capability and more ways to be surprised. The correct amount is the least that does the job.
Conclusion
You can now handle any AI agent claim without needing a definition, because there isn't a settled one. Ask what the vendor means by the term, then ask to see two runs of the same task. If both took the identical route, the software is following a path somebody wrote, whatever it is called. If they diverged because it decided differently, it is directing itself. Then ask the harder question, which is whether your process needs that at all, because the lab that builds these tools recommends starting simpler. If you want the wider picture of deploying and measuring one, our guide to AI agents for business covers it. And if you want that judgement made on your actual process before anyone quotes you a build, talk to Euracle about AI agents.
FAQ
Sources
- Anthropic, Building effective agents, Erik Schluntz and Barry Zhang, December 2024. The acknowledgement that the term is defined in several ways; the agentic systems umbrella; the workflow against agent distinction on predefined code paths versus dynamic self-direction; the description of agents planning and operating independently after a human command; and the advice to start simple.
- OpenAI, defining an agent as systems that independently accomplish tasks on your behalf.
- Model Context Protocol, announced by Anthropic in November 2024 as an open standard letting models connect to external data sources and tools through a unified interface.
