What is agentic AI?

The word marks a change in what software is trusted to do. A chat model writes text and a person decides what to do with it. An agentic system decides some of that itself: it might open a ticket, query a database or send a draft for approval. That moves the risk from bad wording to bad actions, so the questions a buyer asks change from “is the answer good” to “what can it touch, who signs off, and what is recorded”.

The term is also used loosely. Vendors apply it to anything from a single tool call to long-running autonomous systems. IBM defines agentic AI as a system that can accomplish a specific goal with limited supervision. Anthropic groups both fixed workflows and self-directing agents under “agentic systems”. Ask which one a product means before you compare it with another.

Most agentic systems run a loop. The model reads the goal and the current context, proposes the next action, and the runtime executes it. The result goes back into the context and the model decides again. The loop ends when the model says it is done, when a step limit is reached, or when a person stops it.

Tools are what make it agentic. A tool is a function the runtime exposes with a name, a description and an input schema: search a knowledge base, read a record, create an invoice. The model never runs the tool itself. It emits a structured request, and the application decides whether to carry it out. That boundary is where permissions, approvals and logging belong.

Anthropic draws a useful line inside the category. In a workflow, code fixes the path and the model fills in steps. In an agent, the model directs its own process and tool use. Anthropic recommends adding complexity only when it demonstrably improves outcomes, because each extra decision the model makes is another place it can go wrong.

A procurement analyst asks an agent to onboard a new supplier. The agent reads the request, sees that it needs a tax registration and a bank detail check, and calls a document tool to extract both from the attached forms. It then calls a lookup tool against the supplier register to check for a duplicate record.

The duplicate check returns a near match under a different trading name. The agent does not guess. It stops, explains the conflict and routes the case to the analyst. Once the analyst confirms it is a new supplier, the agent drafts the vendor record and pauses again, because creating a payee is an action the team has marked for human approval. The model chose the order of the checks and reacted to an unexpected result; the workflow decided where its authority ended.

Swfte Studio is built to make agents of this kind. You place agents and fixed steps in one visual workflow, give each agent its tools through integrations, and add an approval step that pauses the run until a named person decides. Agents can call any model reachable through Swfte Connect, the OpenAI-compatible gateway in front of many providers.

Swfte Nexus puts policy, approval and audit around coding agents and other agent runtimes, so actions outside an agent’s scope can be blocked or held for a person. Approvals exist in several parts of the platform today and are not one unified inbox. Separation of duties is something you set up through who you assign to each gate; the platform does not enforce it for you.

• Swfte Studio: Build agents and workflows with approval steps
• Swfte Nexus: Policy, approval and audit for agent runtimes

Definitions on this page were read on the sources below on 7 October 2026. Where sources define the term differently, the page says so. The full glossary lists more terms.