Glossary
A working glossary of the terms that come up most often in agentic AI. Updated as the field evolves.
- Agent
- A system that decides which actions to take by combining a model with tools and memory.
- Handoff
- The transfer of control or state from one agent to another, or from an agent to a human.
- Long-horizon task
- A task spanning many steps over hours or days, requiring durable state and memory.
- Model Context Protocol (MCP) MCP
- An open protocol for exposing tools and context to LLMs through a standard interface.
- Multi-agent system
- A system of two or more agents that exchange messages or hand off tasks.
- Planner–critic architecture
- A pattern where a planner proposes steps and a critic prunes or revises them.
- Replay-based evaluation
- Scoring agent candidates against captured real-world sessions with held-out outcomes.
- Retrieval-augmented generation (RAG) RAG
- Retrieving documents at inference time and conditioning generation on them.
- Sandbox
- An isolated execution environment for running agent code or browser actions safely.
- Tool use
- The pattern of an LLM invoking external functions to gather data or take action.