Agents choose their next step from each result
A large language model starts with a goal and tools it has permission to use. It may search, open a file or work through an email connection. The information returned by a tool helps it decide what action would move the task toward completion.
That sequence is what makes the system agentic. IBM describes model outputs being used to accomplish a goal, and Anthropic describes the model directing its own steps and tool use along the way.
- Specify the job. The goal gives the model an outcome to work toward across several actions.
- Make action possible. Tools connect the model to search, files and other systems needed for the task.
- Respond to the finding. What comes back from a tool informs the action the model chooses next.
Chatbots answer, workflows follow paths, agents choose
| Term | What it aims to do | Where it applies |
|---|---|---|
| Generative AI chatbot | Produce content or an answer from a prompt | A conversational exchange |
| Workflow | Run models and tools through steps defined in code | Repeatable jobs with an established sequence |
| AI agent | Choose actions and tools as a model works through a task | Jobs whose next step depends on the result of the previous action |
| Agentic AI | Coordinate agents toward a goal with limited supervision | Systems that carry a job through several actions and results |
IBM describes agentic systems and their relationship to generative AI. Anthropic distinguishes predefined workflows from agents that direct their own steps and tools.
Many products sold as agents are relabeled tools
The label on a product needs to be checked against the job it performs. Gartner's June 2025 assessment counted about 130 real agentic vendors among thousands. It also described agent washing, where assistants, chatbots and automation products acquire an agentic label. For a buyer, a job's files and history provide something to examine beyond that label: the steps it runs, the information it uses and the checks applied to its output.
A team sets goals, tools and approval points
People define the marketing job, connect the systems it needs and review what the agent produces.
Write the job down
Save the goal, inputs and schedule together, so later runs have the same instructions and information to start from.
Limit initial access
Connect a lower-risk system such as the marketing site's repository first, then widen access as experience with the agent grows.
Review outgoing work
Articles, campaigns and site changes wait for approval, while code changes arrive as pull requests for the owner to merge.
Agents add cost, and many projects get dropped
- Examine the project case. Gartner's 2025 forecast put cancellations above 40% of agentic AI projects by the end of 2027.
- Start with simple work. Anthropic advises choosing the simplest suitable setup because agentic systems can require more time and money.
Gartner attributed that forecast to rising costs, unclear business value and inadequate risk controls. Anthropic's 2024 article describes exchanging time and cost for better task performance; the simplest suitable answer may use no agent.