Agentic AI
Agentic AI describes systems that pursue a goal over multiple steps, choosing which tools to call and when to stop, rather than returning a single response to a single prompt. The distinguishing property is control: the model decides the sequence of actions, within limits the deploying organisation sets.
What makes a system agentic rather than generative?
A generative system maps an input to an output. An agentic system is handed an objective and decides for itself what to do first, what to do with the result, and whether it is finished. The sequence is not written in advance by a developer, which is both the capability and the reason the deployment needs limits.
- Goal-directed: the instruction is an outcome, not a function call.
- Multi-step: the number of steps is decided at runtime, not fixed in code.
- Tool-using: progress comes from calling systems, not only from generating text.
- Self-monitoring: the system evaluates its own intermediate results and can change approach.
What are the practical levels of autonomy?
| Level | What the system decides | What it needs |
|---|---|---|
| Assisted | Nothing; it drafts, a person acts | Review before anything happens |
| Supervised | The sequence, with confirmation at each action | A person available in the loop |
| Bounded autonomy | Everything inside a defined tool and budget scope | Tight permissions and complete logging |
| Unsupervised | Everything | Rarely appropriate outside a sandbox |
Most production deployments sit at supervised or bounded autonomy. The level is a policy decision rather than a product feature, and the same software can be configured at any of them, which is why a vendor describing their product as autonomous has not yet answered the question.
What should a buyer ask about an agentic product?
- Which tools can it call, and under whose credentials does each call run?
- Which actions require human confirmation, and can that list be changed by configuration?
- What is logged for each step, and how long is it retained?
- What terminates a run: a step limit, a time limit, a cost limit, or nothing?
- Where does inference happen, and does any internal data leave the organisation to reach it?
Agentic is currently applied to products ranging from a scripted workflow with a language model in one step to genuinely open-ended planning. The answers to these five questions distinguish them; the label does not.
Frequently asked questions
- What is agentic AI?
- Agentic AI describes systems that pursue a goal over multiple steps, choosing which tools to call and when to stop, rather than returning one response to one prompt. The model decides the sequence of actions at runtime, within limits the deploying organisation sets through permissions and confirmations.
- What is the difference between agentic AI and generative AI?
- Generative AI maps an input to an output in a single step. Agentic AI is given an objective and decides the steps itself, calling tools and evaluating intermediate results. Generative describes what the model produces; agentic describes how much of the sequence the system controls.
- Is agentic AI the same as an AI agent?
- An AI agent is the program; agentic is the property that program has. The terms are used almost interchangeably in practice, though agentic AI is more often used to describe a category of system and AI agent to describe a specific deployed one.
- How autonomous should an enterprise agent be?
- Most production deployments run supervised or with bounded autonomy: the agent decides the sequence but irreversible actions need confirmation, tools are scoped narrowly and every step is logged. Full autonomy is rarely appropriate outside a sandbox, because nothing catches an error before it takes effect.
- What questions should a buyer ask a vendor claiming agentic AI?
- Which tools it can call and under whose credentials; which actions require confirmation; what is logged per step and for how long; what terminates a run; and where inference happens. The label covers everything from a scripted workflow to open-ended planning, and these answers separate them.
- Can agentic AI run without cloud services?
- Yes, when the model, the orchestration and the tools all run on local hardware, which is how on-premise and air-gapped agent platforms are built. The constraint is fitting a capable enough model to the available hardware rather than anything inherent to the agentic pattern.