CyberFreezeDevMulti-step admin
Tasks that cross a CRM, a mailbox and an accounting tool before a reply.
Read more about Multi-step admin
Tasks that cross two or three systems, like checking a CRM, a mailbox and an accounting tool before drafting a reply.
- Home
- AI glossary
- AI agent
What is an AI agent?
Unlike a chatbot, an agent acts: it picks tools, checks results and decides the next step.
The plain definition
A language model runs a loop of think, call a tool, read, decide. It never touches your systems; your code runs each call.
Read more about The plain definition
An AI agent is software in which a large language model controls the flow of work. It is given a goal, a set of tools (search, read a file, query a database, send a draft, create a record) and rules. It then runs in a loop: think, call a tool, read the result, decide again.
The tools are exposed through function calling or a standard such as the Model Context Protocol. The model never touches your systems directly; your code runs each tool call and returns the result. For how this differs from a chatbot, see chatbot vs AI agent.
How it works
-
01 Receive a goal
A goal arrives, for example: draft reminders for unpaid invoices over 60 days old.
Read more about Receive a goal
For example: "Find unpaid invoices over 60 days old and draft reminders."
-
02 Plan the next step
The model chooses which tool to call and with what arguments.
Read more about Plan the next step
The model picks a tool and its arguments, such as
list_invoices(status="unpaid", older_than_days=60). -
03 Run the tool
Your code runs the call with the agent's permissions and returns the data.
Read more about Run the tool
Your code executes the call with the agent's permissions and returns the data.
-
04 Observe and repeat
The model reads the result and goes again; step limits and timeouts stop runaway loops.
Read more about Observe and repeat
The model reads the result, then calls the next tool or finishes. A step limit and timeouts stop runaway loops.
-
05 Hand over
Anything irreversible, such as sending, paying or deleting, waits for a person to approve.
Read more about Hand over
Anything irreversible (sending, paying, deleting) goes to a person to approve.
Where it matters in business
Research and preparation
Gathering account history before a sales call, or documents for a compliance check.
Read more about Research and preparation
Gathering account history before a sales call, or collecting documents for a compliance check.
Assistants inside products
Fint's agent proposes records for a person to confirm; Desk's assistant works inside Desk.
Read more about Assistants inside products
In our own products, Fint's AI agent proposes records that a person confirms in an Inbox, and Desk's assistant can be given work to do inside Desk.
A practical example
In Fint, the agent reads a bill and proposes an expense, but only a person can confirm it into the books.
Read more about A practical example
In Fint, an agent can read a bill a user uploads, check whether it is already recorded, and propose an expense. It cannot record money. The proposal sits in an Inbox until a person confirms it. The agent saves the typing and searching; the person keeps control of the books. That split, agent proposes and person confirms, is the pattern Fint is built on, and a sound default for any agent that touches money or customers.
Common mistakes
Too many tools
Forty vague tools lead to wrong picks. Start with a few, precisely described.
Read more about Too many tools
An agent with forty loosely described tools picks the wrong one. Start with a few, named and described precisely.
Broad permissions
Give each agent only the access its job needs, and log every tool call.
Read more about Broad permissions
Give each agent only the access its job needs, and log every tool call.
Agents where a workflow will do
If the steps never change, a fixed workflow is cheaper and more predictable.
Read more about Agents where a workflow will do
If the steps are always the same, a fixed workflow automation is cheaper and more predictable than an agent.
Questions
What is the difference between an AI agent and a chatbot?
A chatbot replies to messages. An agent can take actions through tools and run several steps on its own toward a goal. Many products combine both.
Are AI agents safe to let loose on our systems?
Only with limits: least-privilege access, step limits, logging, guardrails and human approval for irreversible actions. Built that way, an agent is a fast assistant rather than an unsupervised employee.
Can one agent hand work to another?
Yes. That is a multi-agent system, where specialised agents pass tasks and results between them.
Last updated