Terms you'll see throughout this framework. In plain English.
Agent. An AI system that can plan, use tools, act, and observe results, not just generate text. See What is autonomous AI.
Auto mode. A setting (in Claude Code and similar harnesses) that pre-authorizes the agent to run a broad set of actions without asking. Opposite of step-by-step approval. See Permissions.
Claude. Anthropic's family of foundation models. See Claude overview.
Claude Code. Anthropic's official CLI tool for running Claude as an agent in your terminal. See Claude Code overview.
Context window. The amount of text a model can consider at once. Modern frontier models have context windows of 200k–1M tokens.
CoT (chain-of-thought). Technique where the model reasons in text before answering. Improves performance on hard tasks. Related: extended thinking.
Embedding. A vector representation of text. Used in retrieval (RAG) to find semantically similar content.
Evaluation (eval). Testing an agent against a labeled dataset. See Evaluation.
Extended thinking. Claude feature where the model reasons invisibly before producing its answer. See Extended thinking.
Foundation model. A large, general-purpose model trained on broad data. Claude, GPT, Gemini. The "raw" model before any wrapping.
Harness. The runtime wrapping the model, parsing tool calls, executing them, feeding results back. Claude Code is a harness. See Claude Code.
Headless. Running an agent without an interactive terminal or user. Scripts, cron jobs, CI/CD. See Headless.
Hook. A bit of code that runs on an event, tool call pre/post, session start, etc. See Hooks.
LLM. Large Language Model. A model trained to predict the next token.
MCP (Model Context Protocol). An open standard for connecting tools/data to AI models. See What is MCP.
Memory. Persistent state across sessions. See Memory.
Plugin. A packaged extension that adds features (commands, skills, MCPs) to a harness. See Plugins.
Prompt caching. Reusing portions of a prompt across calls to save cost and latency. See Prompt caching.
Prompt injection. An attack where hostile text in a tool result (web page, email, etc.) tricks the model into following unintended instructions. See MCP security.
RAG (Retrieval-Augmented Generation). Pattern where the model fetches relevant documents before answering, instead of relying on training data alone.
ReAct. Reason + Act. A common agent pattern: model reasons, then takes an action, observes the result, reasons again. See ReAct loop.
Reasoning budget. How many "thinking" tokens the model is allowed before answering (extended thinking).
SDK. Software Development Kit, library for calling a model's API.
Skill. A named, slash-command-invokable bundle of instructions for an agent. See Skills.
SSE (Server-Sent Events). A way to stream data from a server to a client over HTTP. Used for MCP remote servers. See Transports.
stdio. Standard input/output. Used for local MCP servers (the client spawns the server as a subprocess). See Transports.
System prompt. Instructions given to the model about its role, rules, and goals, separate from the user's message.
Token. The unit a model sees. Roughly 4 characters in English. Prices and context windows are counted in tokens.
Tool use. Letting the model call functions (tools) that return data. The foundation of agent behavior. See Tool use.
Transport. The communication channel between an MCP client and server, stdio, SSE, or HTTP.