Agent words glossary: tools, memory, context, orchestration and more

Agent documents lean on a small set of terms, and each maker uses them slightly differently. These are the common ones, defined as the makers' own pages use them. The list runs from the basic pieces to the safety words.

The basic pieces

LLM. Short for large language model, the kind of AI that reads and writes text. Makers' pages use the short form constantly.

Augmented LLM. Anthropic's name for a model enhanced with extras such as retrieval, tools and memory¹. Retrieval means fetching information from outside the model, such as a search.

Tool. A function or service the model can ask for. Anthropic's term, tool use, is also called function calling². A tool call is the model's request to use one. See tool calling explained.

Memory. Stored information that carries over. Anthropic's memory tool is described as storing and retrieving information across conversations in files you control². Where memory is kept, and who can see it, depends on the product.

Context window. The text a model can refer to while it writes a response, the response included³. Anthropic adds that as token count grows, accuracy and recall degrade³.

How the pieces are arranged

Agentic system. Anthropic's umbrella term for every variation, from scripted to fully autonomous¹. See what is an AI agent.

Workflow. A system where models and tools are orchestrated through predefined code paths¹. The path is fixed. Agent or fixed workflow compares the two.

Orchestration. Coordinating the models and tools so that the right one acts at the right moment. In Anthropic's orchestrator-workers pattern, a central model breaks tasks down, hands them to worker models, and combines their results¹.

Agent loop. The repeating cycle of plan, act and check. How the agent loop works walks through it.

MCP. The Model Context Protocol, a shared open-source standard that lets AI applications link up with outside systems⁴.

The safety words

Prompt injection. Text from an outside source that carries hidden orders. OpenAI calls it a common and dangerous type of attack, where the outside text tries to override the AI's own instructions⁵.

Guardrails. Checks placed around the model. OpenAI mentions built-in ones that redact personally identifiable information and detect jailbreak attempts⁵.

Tool approvals. A step where a person reviews and confirms an operation. OpenAI's guide says to always enable them for MCP tools, reads and writes included⁵.

Sandbox. A walled-off space for running code. Anthropic's code execution tool runs Python and Bash code in a sandboxed container². Access limits get their own page, permissions and approvals.

Where the words differ

Makers don't share a dictionary. A term in one company's pages may be a product name in another's, and the same idea can carry different labels. When a word matters to a decision, find its definition in that product's own documentation.