AI Agent
Ask a language model a question and you get words back. Give the same model a goal, a set of tools and permission to keep going, and you get something different in kind. An AI agent is a system built around a model that pursues an objective across many steps, deciding for itself what to do next.
What the agent layer adds
The model supplies language and judgement. The agent supplies everything around it:
- A loop — plan, act, observe, repeat — so the result of one step shapes the next.
- Tools: functions, APIs, a shell, a browser. These turn a decision into an effect.
- Memory that outlives a single call, so facts and progress carry from step to step.
- Termination: a goal test, a step limit, a spend cap — something that decides when to stop.
None of this needs a bigger model; it is control flow wrapped around an existing one. ReAct set the basic pattern of interleaving a reasoning step with an action and feeding the observation back, and frameworks such as LangGraph and CrewAI implement that idea with scheduling, retries and tracing attached.
What changes once it acts
The autonomy is both the whole value and the whole problem. An agent can work a task through while you sleep. It can also retry a failing call two hundred times, obey an instruction hidden in a web page it read, or delete the wrong file with total confidence. Everything that separates agent engineering from prompt engineering — permission scoping, sandboxes, approval gates, tracing, budgets — exists because nobody is checking each step.
An AI agent is a system that pursues a goal by repeatedly deciding its own next step — reasoning over the current state, calling tools to act on an environment, and observing the result — until the task is complete. A language model supplies the decisions; the surrounding loop, tools, and memory supply the ability to act. The agent is what turns a model's output into completed work.
A chatbot answers. An AI agent does.
Give it a goal — "find the cheapest flight next Tuesday and book it" — and it works out the steps itself. It searches, reads the results, notices the first option has sold out, tries another, and keeps going until the job is done or it is properly stuck. Along the way it uses real tools: websites, files, other programs.
That is the appeal and the catch. Nobody is watching each step, so these systems are given firm limits — how much they may spend, what they are allowed to touch, and when they must stop and ask a person.