Glossary
A reference for terms used across agentic tooling and this documentation.
The large language model (LLM) that generates responses and decisions — for example, Claude Sonnet or GPT-4. The model reads context (prompt + history + tools) and produces text or tool calls. It has no persistent memory between sessions on its own.
A model augmented with tools and instructions that can take actions — read files, run commands, call APIs. The model alone is passive; the agent acts.
Agent harness
Section titled “Agent harness”The application or CLI that wraps the model and makes it an agent — Claude Code, Cursor, Codex CLI, Windsurf. The harness provides the tool set, loads agent configuration, manages the session, and surfaces the interface to the user.
Different harnesses support different features. A skill that works in Claude Code may behave differently in Cursor. See Commands for a harness compatibility example.
Agent runtime
Section titled “Agent runtime”The execution environment the harness operates in during a session — the process, the working directory, the environment variables, the user’s shell. Skills and hooks run inside the agent runtime.
Session
Section titled “Session”One continuous conversation between a user and an agent harness. Sessions start fresh — the model has no memory of prior sessions unless context is explicitly injected (via AGENTS.md, hooks, or memory files). A session ends when the conversation is closed or reset.
Context window
Section titled “Context window”The total amount of text (measured in tokens) the model can see at once — the conversation history, loaded instructions, tool results, and current message. When the window fills, older content is summarized or dropped. Keeping agent configuration concise preserves space for actual work.
The unit the model uses to process text. Roughly one word or one punctuation mark, but varies by model. Relevant for understanding context window limits and cost.
A function the agent can call during a session — read a file, run a bash command, search the web, write to disk. The harness defines which tools are available. Agent behavior is shaped heavily by what tools are exposed.
Signal
Section titled “Signal”An observable event or output that indicates something happened — a hook firing, a tool call completing, a test passing or failing, a CLI exit code. Skills and disciplines often describe what signals to watch for and how to respond.
A shell command registered to fire when a specific agent event occurs — session start, before a tool call, after a response. Hooks inject always-on behavior without requiring the model to remember to do it. See Commit Discipline for a concrete example.
A reusable, on-demand workflow the agent loads when it matches a situation. Defined in a SKILL.md file. Auto-invoked by the model when the description matches, or explicitly invoked by the user via /name. See Skills.
Discipline
Section titled “Discipline”A class of always-on behavioral rules that shape how any agent (main or sub) operates — not loaded on demand, always active via any channel (hooks, AGENTS.md, or agent definitions). Commit Discipline is an example: rules about when and how to commit that apply in every session.
Discipline and Governance are verified the same way: give the agent a scenario and observe what it does. The distinction is not how you test them — it is when the rule is active (always vs on demand) and what kind of content it encodes (cross-cutting habits vs domain-specific normative rules). See Disciplines.
Subagent
Section titled “Subagent”A specialized agent spawned by the main agent to handle a sub-task — a code reviewer, a judge, a researcher. The main agent delegates, collects the result, and continues. Subagents start fresh with no memory of the parent session unless explicitly briefed.
Prompt
Section titled “Prompt”The text input sent to the model at each turn — the user’s message combined with system instructions, loaded skills, tool results, and conversation history. Everything the model sees is part of the prompt.
Governance
Section titled “Governance”A versioned, agent-readable rule set that tells agents what to do for a specific domain. Loaded on demand via CLI. Dense and imperative — no rationale, no background.
Governance and Discipline are verified the same way: give the agent a scenario and observe what it does. The distinction is when the rule is active (on demand vs always-on) and what kind of content it encodes (domain-specific normative rules vs cross-cutting habits). Governance’s crisp rule text additionally enables static artifact analysis — tools like audit-skill can check documents against governance rules without running an agent. See Governances.
Architecture Decision Record. A frozen document that records why a decision was made — context, trade-offs, rejected options. Written for humans, not agents. See ADRs.