What is an AI agent?
Buy an agent when you needed a workflow and you pay for a model call on every run, plus an outcome nobody can reproduce twice.
AI agent
agentic AI
An AI agent is a program that uses a language model to decide which action to take next, calls tools to take that action, then feeds the result back to the model and repeats until the task is finished or a stop rule fires. The deciding is what makes it an agent: a workflow follows steps a developer wrote in advance.
Three parts have to be present. A model that reasons over the current state. Tools it can call, each with a schema. And a loop that runs until something says stop.
Drop the loop and you have a single model call. Drop the tools and you have a chatbot: it can describe the refund but it cannot issue one. Drop the deciding and you are back to a workflow, which is often the correct answer.
The word decides your failure mode
A workflow fails loudly. Step four throws, the run stops, someone gets paged. You can read the code and know what it would have done.
An agent fails quietly. It picks a defensible wrong action, the run completes, and the log shows a success. Nobody notices until a customer does. That single difference drives everything you have to build around it: permissions per tool, a stop rule, a log of what it chose and why, and a human on the exceptions.
Which is why the honest test is not whether the demo worked. Run the same input twice. An agent can answer differently both times, and your business has to survive that.
- GoalA task, plus what done means.
- DecideModel picks the next action.
- CallOne tool, with typed arguments.
- ObserveResult goes back into context.
- StopDone, budget spent, or escalate.
Take out the arrow from Observe back to Decide and this is a workflow. That arrow is the whole definition, and it is also where the cost lives.
Agent or workflow
Same job, two shapes. Pick by how much the input varies, not by which sounds more advanced.
| Criterion | Workflow | Agent |
|---|---|---|
| Who picks the next step | A developer, in advance, in code. | The model, at run time, from context. |
| Same input twice | Same output. Always. | Usually the same. Not guaranteed. |
| Handles a new input shape | No. It breaks or skips. | Often yes. That is the reason to pay for one. |
| Cost per run | Compute only. | A model call for every decision in the loop. |
| How it fails | Loudly. An exception with a stack trace. | Quietly. A plausible wrong action, logged as success. |
| What you audit | The code. | The trace of what it chose, per run. |
Related questions
01What is the difference between an AI agent and a chatbot?
A chatbot produces text; an AI agent takes actions in your systems. The dividing line is tool access. Ask whether it can create the ticket or only tell you how to create one.
02Do AI agents need a large model?
No, and the deciding step is where model size actually earns its cost. A smaller model is often correct for the tool calls themselves once the schema is tight, so many production agents route between models rather than sending everything to the largest one.
03Are AI agents reliable enough for production work?
Yes, for work where a wrong action is recoverable and a human sees the exceptions. Agents are a poor fit where an action is irreversible and unattended, such as moving money or deleting records, unless a person approves that specific step.
04What is a multi-agent system?
A multi-agent system splits a task across several agents that each own a narrower job and pass work between them. It buys clearer tool permissions per agent and costs you a harder debugging problem, because a wrong answer now has more places to have come from.
