Key facts

Use before
Writing specs, prompts, or review criteria.
Most confused pair
State vs memory.
Highest-risk term
Tool, because it carries permission.
Spec rule
Define terms before granting autonomy.

Terms

Agent

An application that can plan, call tools, keep task state, and complete multi-step work with some autonomy.

Agent loop

The repeat cycle where a model selects an action, the application executes it, the result returns, and the model decides what to do next.

Tool

A typed capability exposed to the model, such as search, database lookup, file write, code execution, payment, or browser control.

Workflow

A controlled sequence of steps. A workflow can use LLM calls without giving the model broad autonomy.

Memory

State carried across turns or runs. Short-term memory keeps a thread coherent; long-term memory stores durable facts or preferences.

Checkpoint

A saved execution state that lets an agent resume after an interruption, review, crash, or long pause.

Guardrail

A policy or validation layer that checks user input, tool calls, retrieved context, or final output.

HITL

Human in the loop. A deliberate approval, review, or escalation point before the agent takes a sensitive step.

Handoff

A controlled delegation from one agent or specialist to another.

MCP

Model Context Protocol, a standard for connecting AI applications to external resources, prompts, and tools.

Trace

A structured record of model calls, tool calls, decisions, guardrails, cost, latency, and outputs during an agent run.

Evaluation

A test suite that measures whether the agent completes tasks correctly, safely, cheaply, and consistently.

Excessive agency

A security failure mode where a model is allowed to perform damaging actions because its permissions or tools are too broad.

Tool schema

The argument contract a model must satisfy before your application runs a tool.

Deterministic step

A step implemented in normal code because the desired behavior is known and should not be left to model judgment.

How to use this glossary in a spec

Pick the vocabulary before implementation starts. If the team says "memory," clarify whether that means current conversation state, a saved user preference, a vector store, a database record, or a resumable checkpoint. If the team says "tool," clarify whether it is read-only, writes internal state, contacts a customer, spends money, or calls another agent.

The purpose is not academic precision. The purpose is to remove ambiguity before a model receives permissions. In agent work, unclear words become unclear boundaries.

Sources used

  1. Agents SDK OpenAI Developer Docs. Accessed 2026-07-06.

    Defines agents as applications that plan, call tools, collaborate, and keep state.

  2. Model Context Protocol Specification Model Context Protocol. Accessed 2026-07-06.

    Official specification for resources, prompts, and tools exposed by MCP servers.

  3. Agents CrewAI. Accessed 2026-07-06.

    Defines CrewAI agents as autonomous units with roles, goals, tools, memory, and collaboration.

  4. Persistence LangChain Docs. Accessed 2026-07-06.

    Explains short-term memory through checkpointers and long-term memory through stores.