Agentic AI means the model chooses the next step
A chatbot answers and stops. An agentic system is given an objective, decides what to do, does it, observes the result and decides again. The loop is the whole definition.
The four questions that separate agentic from marketing
- Does it choose its own actions, or follow a fixed path?
- This is the load-bearing test. A workflow with a model at one node is a workflow. If the sequence of tool calls is determined before the run starts, the system is not agentic no matter what the product page says.
- Does it observe the result and change course?
- Acting without looking is a script with extra steps. The feedback edge is what makes the loop a loop, and it is the part most often missing from systems described as agentic.
- Can it affect something outside itself?
- Reasoning that produces only text has no blast radius. As soon as it can send, buy, file or delete, every question about authorisation, reversibility and audit becomes real and urgent.
- Does it decide when it is finished?
- The stopping condition is where cost lives. A system that cannot recognise completion runs until it hits a turn limit, and that limit is the actual budget whether or not anyone set it deliberately.
Three things the word gets used for
| What it is | Agentic by the test above | |
|---|---|---|
| Chatbot with retrieval | Fetches context, then answers once | No, the path is fixed |
| Workflow with an LLM node | Deterministic graph, model at one step | No, the graph decides |
| Tool-calling loop | Model picks tools until it is done | Yes, this is the base case |
| Multi-agent system | Several loops, one delegating to others | Yes, with more failure modes |
| Browser agent | A loop whose tools act on a live page | Yes, and the blast radius is external |
| Scheduled script calling an API | Cron with a model somewhere in it | No, regardless of the label |
As of 2026-08-31 · Competitor details come from each vendor's published pricing page on that date.
Common questions
Is agentic AI the same as an AI agent?
Agentic AI is the property, an AI agent is a thing that has it. The adjective is useful precisely because it is a matter of degree: a system can be somewhat agentic, with a model choosing among three tools inside an otherwise fixed pipeline.
Why does everything suddenly claim to be agentic?
Because the word has commercial value and no agreed definition, which is a combination that always produces inflation. The four questions above are worth asking of any product using the term, and most answers are no to at least two of them.
What actually makes these systems hard to operate?
Non-determinism plus external effects. Either alone is manageable. Together they mean the same input can produce a different irreversible action, which breaks the testing and rollback assumptions most engineering practice is built on.
Where should I put the guardrails?
Outside the model, on the tools. Instructions in a prompt are a request; a permission check in the tool implementation is a control. Anything that must not happen has to be impossible rather than discouraged.
A bounded place for the loop to act
Isolated sessions, credentials the model never sees, and a record of every action taken.