Harness, MCP, ACP, compaction, YOLO mode: the words that decide what your coding agent is allowed to do are the ones most pages skip. This AI coding glossary defines each in a sentence or two, and links the ones that have a full guide.
A–C
ACP (Agent Client Protocol): an open standard that lets a coding agent run inside any editor that supports it, such as Zed or JetBrains IDEs. Not the same as IBM’s Agent Communication Protocol. Guide
Agent (AI coding agent): a program that uses a language model to do a task in your project: it reads files, runs commands, edits code and checks its own work in a loop until the task is done. Guide
AGENTS.md: a Markdown file in a repository that tells coding agents how to build, test and work on the project. A README for agents. Guide
Approval prompt: the moment an agent stops and asks you before doing something your rules say needs a human, such as running a command or editing a file. Guide
Benchmark, leaderboard: a fixed set of coding tasks used to rank models or agents by how many they finish, such as SWE-bench Pro or Terminal-Bench. Good for a shortlist, not for picking a winner. Guide
Chunking: splitting documents into passages before embedding them, so each vector stands for one idea instead of a whole manual. Guide
Context window: how much text a model can take in at once, measured in tokens. For an agent, it holds the instructions, tool definitions, files it read, command output and the conversation so far.
Context rot: the drop in accuracy as a context fills up. Long sessions get worse, not only more expensive. Guide
Cosine similarity: the usual measure of how close two embeddings are in meaning: how nearly they point in the same direction. Guide
E–K
Embedding: a list of numbers that represents what a text means, so texts with similar meanings get similar numbers. Guide
Flash attention: a way of computing attention that uses less memory as the context grows. Local runtimes such as Ollama turn it on where the hardware supports it. Guide
Function calling, tool calling: a model’s ability to answer with a structured request to run a tool, such as “read this file”, instead of plain text. Every coding agent is built on it. Guide
Grouped-query attention (GQA): a model design where several attention heads share one set of keys and values, which shrinks the KV cache. The reason you size memory from KV heads, not attention heads. Guide
Harness: the program around a model that turns it into a coding agent: it runs the loop, offers the tools, builds the context and enforces permissions. Claude Code, Codex and OpenCode are harnesses. Guide
HNSW: the most common index in vector databases: a graph that links each vector to its neighbors, so a search hops toward the closest ones instead of checking everything. Guide
KV cache: the memory a model keeps of the keys and values for every token it has read, so it doesn’t recompute them. It grows with the context and is why long contexts need so much GPU memory. Guide
L–P
Local model: a language model that runs on your own hardware, for example through Ollama or vLLM, so prompts and code never leave your machine. Guide
MCP (Model Context Protocol): an open standard for giving AI apps tools and data. A service is wrapped once as an MCP server and works in every app that supports MCP. Guide
Prefill: the step where a model reads the whole prompt before it writes anything. For long prompts it is most of the work, and it is what prompt caching skips. Guide
Prompt caching: reusing the work done on the beginning of a prompt when the next request starts the same way, which makes it faster and cheaper. Agent sessions depend on it. Guide
Prompt injection: instructions hidden in text the model reads, such as a web page, an issue comment or a document, that try to make it do something you didn’t ask for. Guide
Permission rules: rules the agent’s app checks before every action: allow, ask first or block, for reading files, editing them, running commands and calling tools. Guide
Q–S
Quantization: storing a model’s numbers, or its KV cache, with fewer bits, such as 4 or 8 instead of 16. Smaller and faster, with some loss of precision. Guide
RAG (retrieval-augmented generation): looking up the passages a question needs before the model answers, and giving them to the model to answer from. Guide
Reasoning tokens, thinking: tokens a model generates to work through a problem before answering. They are billed as output, even when you only see a summary. Guide
Sandbox: a boundary around what an agent’s commands can reach, such as a container, a VM or an operating-system restriction. Unlike permission rules, it also stops scripts the agent writes. Guide
Semantic search: search by meaning instead of exact words, usually built on embeddings. Guide
T–Z
Terminal coding agent: a coding agent you run as a command in your project folder, such as Claude Code, Codex CLI or OpenCode. Guide
Token: the unit models read and write: a word, part of a word or a symbol. Context sizes and prices are counted in tokens.
TUI (terminal user interface): a full-screen, keyboard-driven app inside a terminal. Most terminal coding agents are TUIs. Guide
Vector database: a database that stores embeddings and finds the ones closest to a query quickly. Guide
Worktree (git worktree): a second working folder for the same repository, on its own branch. The way to run two agents on one project without them overwriting each other. Guide
YOLO mode: running a coding agent with its approval prompts turned off, so it acts without asking. Each tool has its own switch, and they don’t all turn off the same things. Guide