AI agent A model that decides the next step and acts, instead of only answering.
Short answer. An AI agent is a system in which a language model decides what to do next and uses tools to do it, repeating until the task is done or a limit is reached. OpenAI describes agents as systems that independently accomplish tasks on your behalf. Anthropic separates agents, where the model directs its own process and tool use, from workflows, where code fixes the path. In an office, a chatbot answers “what is our leave policy?”. An agent can read the list of unpaid invoices, draft a reminder for each client, and wait for a person to approve the emails before any is sent. The more an agent can do, the more its permissions and limits matter.
01
How an agent works
An agent runs in a loop. The model reads the request and the list of tools it may use, and either answers or asks for a tool. The application runs the tool and returns the result, and the model decides again. Anthropic describes agents as typically just models using tools, based on feedback from their environment, in a loop.
The loop ends in one of three ways: the model judges the task complete, the agent fails and hands control back to the user, or a stopping condition fires. Anthropic notes that it is common to include stopping conditions such as a maximum number of iterations, and OpenAI’s guide says that on failure an agent can halt execution and transfer control back to the user.
02
Agent, chatbot or workflow?
The words are used loosely. This table follows the definitions in the OpenAI and Anthropic guides listed below. OpenAI says simple chatbots, single-turn model calls and sentiment classifiers are not agents, because they do not use the model to control how a task runs.
Kind
Who chooses the steps
Office example
Chatbot or single model call
Nobody: the model answers one prompt
Ask for the notice period in a staff contract
Workflow
Your code fixes the path; the model handles parts of it
Read each GST invoice, extract the fields, post them to the ledger
Agent
The model picks the steps and the tools
Chase unsigned vendor forms: read the tracker, draft reminders, ask for approval
03
What can go wrong
OWASP calls the main risk in agent systems excessive agency: damaging actions taken in response to unexpected, ambiguous or manipulated model output. Triggers include a model that makes things up and prompt injection, where instructions hidden in an email or web page steer the model. The root causes OWASP lists are excessive functionality, excessive permissions and excessive autonomy.
OWASP’s own example is an assistant given a mailbox so that it can summarise mail, through a plug-in that can also send mail. A crafted incoming email tells it to forward the inbox to an outsider. A plug-in that can only read, a read-only permission, or a person pressing send would each have stopped it. Anthropic adds that autonomy brings higher costs and the potential for compounding errors.
04
Before you give an agent a job
Name one task and one person who owns it.
Start with read-only tools, and add tools that change things one at a time.
Run the agent as the user who asked, not as an administrator, so it cannot see more than that person can.
Cap the number of steps and the number of retries.
Require a person’s approval for anything that sends, pays, deletes or changes access.
Log every model call and tool step with the name of the person who asked, so a bad outcome can be reviewed.
Test on twenty real examples from your own work before staff rely on it.
05
Where ANVE stands
ANVE’s agent loop allows up to seven tool steps per request, and a claim made without a successful tool call is challenged. ANVE’s security page (/security) says the tools an agent can use must be approved by an administrator, a tool that changes is switched off, and every tool call is logged. Agents for legal research, compliance checks, tender responses, action follow-up, a policy helpdesk and a voice front desk built on AnveVoice are built with a paid pilot. They are not shipped products, and ANVE itself is pre-release.
06
Questions people ask
Is a chatbot an AI agent?
Not by OpenAI’s definition. Simple chatbots and single-turn model calls do not use the model to control how a task runs, so they are not agents.
Does an AI agent need the internet?
No. An agent needs a model and tools. If both run on your own network, nothing has to leave it. Tools that search the web or call a cloud model do need a connection.
How many steps should an agent be allowed?
As few as the task needs. Anthropic notes that it is common to set a stopping condition such as a maximum number of iterations. ANVE’s loop stops at seven tool steps per request.
Can an agent act without approval?
It can if you allow it. OWASP and OpenAI both advise requiring human approval for high-impact or irreversible actions such as payments, deletions and outgoing messages.
Which tools may an ANVE agent use?
Only tools an administrator has approved. ANVE’s published security page says a tool that changes is switched off and every tool call is logged. ANVE is pre-release.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
