A plain definition
A chat assistant answers questions. An AI agent goes further: given a goal, it decides what steps to take, uses tools to take them, looks at the results and continues until the task is done or it needs help.
Ask a chat assistant “which clients still owe us records?” and it can only answer from what you paste in. Ask an agent the same question and it might search your practice system, check your document store, compile a list and draft reminder emails.
What makes something an agent
There is no single agreed definition, but agents usually have:
- A model that reasons about what to do next.
- Tools it can call, such as search, read a file, create a draft, run a query.
- A loop: act, observe the result, decide the next step.
- Instructions or a goal that frame the work.
- Some memory or state, whether within one task or across tasks.
The same AI product can behave as a chat assistant in one moment and an agent in the next, depending on which tools are connected. Claude, ChatGPT, Copilot and coding agents such as Codex can all act as agents when given tools.
How agents reach business systems
Agents use tools through connectors. A common standard for this is the Model Context Protocol, which lets an AI application connect to “servers” that expose tools and data. Other routes include vendor-specific plug-ins and direct API integrations.
Why agents change the risk
When AI only produces text, a person reads it and decides what to do. When AI can act, the decision may happen without a person. That changes the questions an organisation has to answer:
- Who is the agent acting for? Its access should not exceed that person’s.
- What can it see? Only the data the task needs.
- What can it do? Drafting is not sending; reading is not changing.
- When must a person decide? Consequential actions should wait for approval.
- What happened? Every action should be recorded.
Common kinds of business agent
- Assistants with tools, used by one person for their own work.
- Workflow agents that run a defined process, such as preparing a monthly pack.
- Coding agents that read and change software repositories.
- Background agents that run on a schedule or in response to events.
- External agents used by clients, suppliers or partners that interact with your organisation.
What agents are not
Agents are not infallible, and they are not accountable. They can misread instructions, pick the wrong tool or be misled by content they read. Responsibility for what an agent does stays with the organisation and the person it acts for.
How Kroy approaches it
Kroy treats each agent as something to identify and govern, not a black box with a key. When an agent makes a request, Kroy checks the person it acts for, the agent itself, what the person has delegated to it, your policies and whether approval is required. Skills give agents defined jobs with defined permissions.