Glossary · Category

Agents & Tool Use

8 plain-English definitions.

Agentic loop

The cycle of an LLM-based agent perceiving state, reasoning, taking an action (often a tool call), and observing the result, repeated until a task completes.

Computer use / GUI agent

A model capability to perceive a screen (via screenshots) and control a computer by generating mouse/keyboard actions to complete tasks.

Function calling / tool calling

A model capability where the LLM outputs a structured call to an external function or API (with arguments) instead of, or alongside, natural-language text.

Multi-agent orchestration

Architectures where multiple LLM agents with different roles or specializations collaborate, delegate, or critique each other to complete complex tasks.

Planner-executor pattern

An agent architecture that separates a high-level planning model (deciding what to do) from an execution model or tool layer (doing it).

ReAct (Reason + Act)

A prompting framework that interleaves explicit reasoning steps with tool-use actions, letting a model plan, act, and observe results iteratively.

Self-reflection / self-critique

An agentic technique where a model reviews and critiques its own prior output or reasoning to catch errors before finalizing an answer.

Tool schema

The structured (typically JSON Schema) description of a tool's name, parameters, and expected input format that a model uses to generate valid tool calls.