Agent teams
Several separate AI sessions that work together. One session acts as the team lead, and they share a task list and can message each other. Unlike a subagent, each teammate has its own context window.
Plain-language definitions for agentic AI and coding-agent development. Reuse them in your own project.
Showing 1–20 of 46 terms
glossary.md: 46 terms, with sources
Several separate AI sessions that work together. One session acts as the team lead, and they share a task list and can message each other. Unlike a subagent, each teammate has its own context window.
A way of working where the AI reads files, runs commands, and edits code by itself while you watch or step in. A chat assistant only replies with text, and you apply it yourself.
Everything around the model that turns it into a coding agent: file access, running commands, permission checks, memory loading, and the loop that chains steps together. The harness is the tool you run. The model is the AI inside it.
A markdown file where you write instructions for AI coding agents. If a project has one and no CLAUDE.md, Claude Code reads it as the project instructions.
A live web page that the agent publishes from your session to a private link, so you can look at the result visually or share it instead of reading terminal text. The page updates in place when the session republishes it.
Notes the agent writes for itself from your corrections and preferences. They are saved per repository and loaded at the start of each session. It is the agent-written counterpart of CLAUDE.md, which you write yourself.
A permission mode where a separate classifier model reviews actions instead of you. Most actions run without asking, while risky ones, like going beyond the task or using untrusted infrastructure, are blocked.
A start-up option that skips hooks, skills, plugins, MCP servers, memory, and CLAUDE.md. It gives the same result on every machine, which suits CI and scripts.
Ready-made skills that ship with the tool, such as /batch, /debug, and /loop. Instead of running fixed logic, they give the agent a detailed prompt and let it plan the work.
A restore point saved at each prompt you send. The tool snapshots files before every edit, so you can roll code, the conversation, or both back to an earlier point. It is separate from git and does not track changes made through shell commands.
The folder where the agent reads project settings: settings, hooks, skills, subagents, rules, and memory. A project keeps one at its root, and your personal defaults live in a folder in your home directory.
A markdown file of standing instructions you write for the agent. It is loaded at the start of every session. Put project conventions, architecture notes, and "always do X" rules here. It survives compaction. The agent can also read AGENTS.md.
A session that keeps running after you close your laptop because it runs on cloud servers instead of your machine. A session in your own terminal or editor is a local session.
A reusable instruction you run by typing /name. Built-in commands such as /clear and /compact control the session, and you can add your own. Skills are the recommended way to package multi-step commands.
Automatic summarizing of the conversation when the context window is nearly full. Old tool outputs are cleared first, then the conversation is summarized. Instructions that only exist in the chat may be lost, but CLAUDE.md is reloaded.
The working memory of a session. It holds the conversation, file contents, command output, CLAUDE.md, loaded skills, and system instructions. As it fills up, compaction summarizes it.
A setting that controls how much the model thinks on each step. Higher effort means more thinking and deeper reasoning. Lower effort is faster and cheaper.
Step-by-step reasoning the model does before it answers. You can tune it with the effort level. The thinking is shown in the terminal in gray text.
A handler you define that runs automatically at a fixed point in the agent lifecycle, such as before a tool runs, after a file edit, or at session start. A hook can run a shell command, call a web endpoint, or run another agent. Hooks are deterministic: they fire at set points, not at the model discretion.