What is an AI agent?
An AI agent is a system that uses a language model to choose and perform actions, rather than only producing text. The distinguishing feature is not intelligence but consequence: an agent can read from and write to systems outside the conversation.
A chatbot answers. An agent acts. The moment a model can call a function that changes something — create a ticket, update a record, issue a refund — the engineering problem stops being about prompt quality and becomes about authority: what is this thing allowed to do, on whose behalf, and who answers for it afterwards.
That shift is why most agent projects stall between a working demo and production. The demo needs a model and a function. Production needs identity, scope, approval, isolation and an audit trail, none of which the model provides.
An agent run is bound to one end user inside one host tenant, can call only actions a human reviewed and published, and pauses before anything consequential.
Related terms
- Agentic AI — An umbrella term for AI systems that plan across multiple steps and act on the world, rather than answering one prompt at a time.
- Tool calling — The mechanism by which a language model requests that a specific function be run with specific arguments.
- Human in the loop — A design where a person approves or rejects a consequential action before it executes.
See it working rather than described.
The product demo runs one agent across three systems and stops at the approval gate.