Strip away the marketing and an AI agent is three things bolted together: a model that decides, tools that act, and a loop that keeps deciding and acting until a goal is met. The model alone is not an agent — it only produces text. The automation alone is not an agent — it only follows rules. The agent is the combination: software that can be given an outcome rather than a procedure, and that holds real credentials to real systems while it pursues it.
That definition does useful work because the word is applied to almost everything right now. A scripted chatbot, a coding assistant, and an autonomous procurement workflow are all sold as "AI agents," and the label tells you nothing about which one can spend money. The questions that classify a system are operational: does it carry a goal across steps without re-prompting, does it write to anything, and how far does its loop run unattended? The agentic AI pillar covers the quality those questions measure; the agents-vs-chatbots and agents-vs-agentic comparisons draw the lines in detail.
The types you will actually meet
Classic AI theory sorts agents by sophistication — reflex agents that map condition to action, model-based agents that track state, goal-based and utility-based agents that plan. The taxonomy that matters operationally is simpler and crosses those lines: what does it do, and what can it touch. Coding agents open pull requests. Support agents read and write customer records. Voice agents make spoken commitments on calls. Back-office agents reconcile systems of record. Multi-agent systems split one workflow across several specialists. And no-code agents are any of the above, assembled by whoever had an afternoon and a license — which is why they dominate the unregistered population.
Whatever the type, the anatomy repeats: model, tools, state, orchestration — the architecture page walks it properly. The tool list, not the model, is the agent's real capability boundary, and that single fact organises most of what follows.
When an agent becomes your problem
An agent in production is an actor inside your organisation: it holds credentials, takes actions at machine speed, and works around the clock without supervision unless you built the supervision in. Readiness for that actor is concrete, sequenced work — an inventory of what already runs, an identity and scoped permissions per agent, security against the attacks agents specifically attract, governance that decides who may approve and widen them, and autonomy granted on evidence rather than enthusiasm. Teams meet this work either before their first consequential agent or during their first agent incident; the content of the work is identical, the price is not.