AI Agents

An AI agent is a system in which an LLM is given a goal and a set of tools, then loops through calling those tools, observing the results, and deciding what to do next, until the goal is reached or a stopping condition is hit. The defining feature is that the model controls its own next action, rather than producing one response in one call.

How it works

An agent loop typically goes: the model is shown the goal and the available tools; it picks a tool and produces a call; the runtime executes that call; the result is fed back into the model; the loop continues until the model signals completion. Tools commonly include web search, code execution, file access, retrieval over an internal corpus, and other LLM calls.

Common patterns

• ReAct. Interleaves Reasoning text with Acting tool calls. The original pattern from the 2022 paper.
• Tool calling. Models trained to emit structured tool-call JSON natively, with the runtime executing and returning results.
• MCP (Model Context Protocol). A standard interface between agents and tool servers, allowing tools to be shared across agent frameworks.
• Multi-agent systems. Multiple agents collaborate, debate, or specialise across subtasks.