Agentic AI Glossary
Definitions for the terms that come up when you build and run AI agents in production.
32 terms
Foundations
Agent2Agent Protocol (A2A) The Agent2Agent (A2A) Protocol is an open standard that lets independent AI agents discover each other, exchange tasks, and collaborate as peers. Agentic AI Agentic AI describes systems that pursue a goal autonomously, reasoning and calling tools across many steps instead of replying once. Agentic Workflow An agentic workflow is a multi-step process in which AI agents decide some steps at runtime, mixing model-driven choices with set structure. AI Agent An AI agent pairs a language model with tools, memory, and a loop so it can take actions toward a goal, not only generate text. Memory Agent memory carries information across an agent's steps and sessions, so it retains context beyond a single model call's window. Model Context Protocol (MCP) MCP is an open protocol that lets AI agents connect to tools and data sources through one standard interface instead of bespoke integrations. Multi-Agent System A multi-agent system splits a task across several specialized AI agents that coordinate, by handoffs or a supervisor, toward one goal. Retrieval-Augmented Generation (RAG) RAG grounds a language model's output by retrieving relevant documents at query time and placing them in the prompt, not training data alone. Structured Output Structured output constrains a language model to emit data in a defined schema, such as JSON, so downstream code can parse results reliably. Tool Use (Function Calling) Tool use lets a language model call external functions — APIs, code, databases — by emitting a structured request the runtime then executes.
Durability, observability & control
Approval Gate An approval gate is a checkpoint where an agent pauses for human sign-off before a sensitive action runs, holding state until someone responds. Crash Recovery Crash recovery lets an agent resume from its last completed step after a process dies, restoring durable state instead of restarting from scratch. Determinism & Replay Determinism and replay let a runtime reconstruct an agent run from recorded step results instead of re-executing model calls or side effects. Durable Execution Durable execution persists an agent's state server-side so a run survives crashes and restarts, resuming from the last completed step instead of starting over. Evaluation (Evals) Agent evaluation measures how well an agent performs by scoring its runs against criteria such as task success, tool-call accuracy, and faithfulness. Execution History (Audit Trail) Execution history is the ordered, persistent record of every step an agent run took, from LLM and tool calls to approvals and retries. Guardrails Guardrails are validation checks on an agent's inputs and outputs that enforce rules at runtime, blocking or correcting unsafe behavior. Human-in-the-Loop (HITL) Human-in-the-loop pauses an agent so a person can review, approve, or correct an action before the run continues on high-stakes steps. Idempotency Idempotency is the property that running an operation more than once has the same effect as running it once, making agent tool retries safe. LLM-as-Judge LLM-as-judge is an evaluation method that uses a language model to score or compare another model's outputs against a rubric at scale. Observability Agent observability is the ability to see what an agent did and why, capturing its steps, tool calls, inputs, outputs, and decisions for debugging. Orchestration Agent orchestration is the layer that coordinates an agent's steps, tool calls, and sub-agents, deciding what runs when and recovering from failures. State Management Agent state management is how a run's execution position and data are stored and updated, so the agent always knows where it is and what it has done. Tracing & Spans Agent tracing records each step of a run as nested, timed spans, producing a structured trace that shows what an agent did, in what order, and why.
Patterns & reasoning
Choreography vs. Orchestration Orchestration coordinates agents from a central controller, while choreography lets agents react to events and coordinate without a conductor. Context Engineering Context engineering is the practice of curating what information fills an agent's context window at each step of a run. Handoff A handoff is when one agent transfers control and context of a task to another agent better suited to continue it, in a multi-agent system. Planning Agentic planning is when an agent breaks a goal into an ordered set of steps before or during execution, then carries them out and revises as needed. ReAct (Reasoning + Acting) ReAct is an agent pattern that interleaves reasoning steps with actions, letting a model think, take a tool action, observe the result, and adjust. Reasoning Loop The agent loop is the cycle in which an agent reasons, picks an action, runs it, observes the result, and repeats until a stop condition is met. Supervisor Agent A supervisor agent coordinates a multi-agent system by decomposing a task, delegating subtasks to worker agents, and integrating their results. Swarm An agent swarm is a multi-agent pattern in which control passes peer to peer via handoff conditions rather than through a central coordinator.