AI agents are programs that use a large language model to plan and carry out multi-step tasks — retrieving information, calling tools and taking actions — towards a goal. Unlike a chatbot, which answers one message, an agent decides what to do next based on what each step returned.
Also called: AI agent, autonomous agents, LLM agents
An AI agent is built from a few parts: a language model that reasons, instructions that define its job and limits, tools it may call (a search over documents, a database query, an API, an e-mail draft), the context it carries between steps, and a stopping condition. On each step the model looks at the goal and everything gathered so far, chooses the next action, and reads what came back. That loop is what separates an agent from a single prompt.
It also separates agents from older automation. Robotic process automation follows a script and breaks when a screen or a document changes. A chatbot answers but does not act. An agent can handle the messy middle — an e-mail written in its own words, a contract with an unusual clause, a request that needs two systems — because it decides the path at run time. The same flexibility is why it needs controls a script never did.
A worked example from banking: complaint handling. A customer e-mail arrives. The agent classifies the complaint type and urgency, looks up the account through a read-only connector, searches the bank's complaints policy in a private knowledge base, and drafts a reply that cites the policy section it relied on. It notes the regulatory response deadline and routes the draft to a case handler, who edits and approves it before anything is sent. Every step — the documents retrieved, the model calls, the approver — is in the run log. The handler reviews instead of researching, and the bank can show afterwards exactly why each reply said what it said.
The common failures are predictable. Agents are given write access they do not need. They are released after a good demo instead of a test set that covers awkward cases. Instructions hidden in an incoming e-mail or document — prompt injection — are treated as commands. A loop that keeps calling a model runs up cost overnight. And nobody can reconstruct what happened when an answer is challenged. Each of these is an engineering and governance problem, not a model problem.
The pattern that works in regulated organisations is bounded autonomy: an agent may read and draft freely, but anything consequential — sending, paying, changing a record, disclosing data — passes a guardrail and a human approver. Versions are immutable, a second person approves promotion to production, budgets are capped and every run is traced.
That is how ATX Agent Hub is built. Agents are assembled in a visual workflow builder, answer from private knowledge bases with citations, run behind AI Guardians, Fact Checking, Intent Check and Domain Constraint guardrails, pause for human-in-the-loop approval where you place it, and are released through four-eyes promotion with role-based access and immutable audit logs.
Frequently asked questions
Are AI agents the same as chatbots?
No. A chatbot answers a message. An AI agent works towards a goal over several steps — retrieving information, calling tools and taking actions — and decides each next step from the results of the last one.
Can AI agents act without human approval?
Technically yes, but in most enterprise settings they should not for consequential actions. The usual design lets an agent read and draft on its own, and requires a person to approve anything that sends, pays, changes a record or discloses data.
What does an AI agent need to run in an enterprise?
Connections to the systems it works with, a knowledge base for grounded answers, guardrails, human approval steps, versioning with independent release, access control, trace logs and cost limits — the parts an AI agent platform provides around the model.
How do you test an AI agent?
With a test set of realistic and awkward inputs and the expected behaviour for each, run before every release and compared with the previous version. Production traces then show where real inputs differ from the test set.
This site needs JavaScript for the full experience. Every page above is reachable without it.