// agentic
An AI agent goes further than a chat: it acts. It combines an LLM, a reasoning loop and executable tools (shell, API, browser) to accomplish multi-step tasks without supervision at every step.
Components
- Planner — breaks the objective into sub-tasks.
- Tools — functions that act on the system.
- Memory — persistent context between steps.
The diagram shows the LLM at the center, connected to the planner then to the tools in a loop. It is powerful — and risky. An agent can write to your disk or call a paid API.
Golden rule: sandboxing is mandatory, least privilege, exhaustive logs.