Coding agent glossary
Reviewed November 2026
Agent mode
Agent mode is a chat mode in which the assistant carries out a task itself, editing files and running steps, instead of only answering. You choose it for work you can describe as a goal and review afterward. In VS Code's Copilot docs, an agent uses a model and tools to find relevant code, make changes, and run checks, while you review the results. Cursor's docs say that for quick or familiar tasks, jumping straight to Agent mode is fine.
Agent Skills
Agent Skills are packaged instructions for a specific task that an agent loads when the task calls for them. Typical examples are a release checklist or a review procedure, which Claude Code's docs list as task instructions that suit a skill. In Claude Code's docs, a skill holds instructions for one kind of task, and Claude loads it only when you invoke it or the task matches. To share project skills, you commit the .claude/skills/ directory to version control.
AGENTS.md
AGENTS.md is a simple, open format for guiding coding agents, written as just standard Markdown. It matters when several agents need the same project instructions. The agents.md site calls it a README for agents that holds build steps, tests, and conventions, and says the closest AGENTS.md to the edited file wins. OpenAI's Codex docs state that Codex reads AGENTS.md files before doing any work.
Allowed tools
Allowed tools are the tools an agent may run without asking a human for approval first. The list matters most when an agent works unattended, as in headless runs. Claude Code sets the list with the --allowedTools flag, whose tools execute without prompting for permission, apart from reads from network paths. Deny rules win over this list: if a tool is denied at any level, no other level can allow it.
Architect mode
Architect mode is a chat mode that gives one coding request to two model roles, one that plans the change and one that writes the file edits. Engineers turn to it when a model reasons well but writes poor edits. In Aider, an architect model will propose changes and an editor model turns that proposal into specific file edits. The docs state the trade-off: it uses two LLM requests, which can take longer and increase costs. Aider's special shortcut --architect launches straight into this mode.
Ask mode
Ask mode is a chat mode for questions about code, in which the assistant answers and explains without changing your files. Engineers use it to understand unfamiliar code or test an approach before handing work to an agent. For a read-first step in Cursor, Plan Mode creates detailed implementation plans before writing any code. The agent researches the codebase, asks clarifying questions, and generates a reviewable plan you can edit before building.
Branch protection
Branch protection is a repository setting that enforces rules on pushes and merges to chosen branches. Teams set it up before agents open pull requests against a main branch. On GitHub, each branch protection rule disables force pushes and prevents matching branches from being deleted by default. Such rules can also stop an agent, since a ruleset or branch protection rule can prevent Copilot from creating or updating pull requests. Where a repository requires pull request approvals, your approval of a Copilot pull request won't count toward the required number.
CLAUDE.md
CLAUDE.md is the file Claude Code's settings docs name for giving Claude standing instructions, alongside a system prompt flag. Its CLI reference says project conventions Claude should always follow belong there, and the commands page describes running /init to generate a starter CLAUDE.md. For a repository already set up for other agents, the memory docs add that /init reads other tools' instruction files and folds the relevant parts into the generated file.
CODEOWNERS
CODEOWNERS is a repository file that maps file patterns to the people or teams responsible for that code. It comes into play when agent-written pull requests touch owned code, a common stop in a review loop. On GitHub, code owners are automatically requested for review when someone opens a pull request that modifies code they own. Repository owners can edit their protection rule and turn on Require review from Code Owners. With that on, an approval from any one of the owners is sufficient.
Content exclusion
Content exclusion is a setting that keeps chosen files or repositories out of what an AI coding assistant may read or use as context. Administrators and repository owners set it up when part of a codebase is sensitive, such as credentials or licensed code. When several agents share one repository, look up each tool's own exclusion setting before you assume a file is hidden from all of them, and test the result with a sample file that should stay out of reach.
Copilot setup steps
Copilot setup steps are a GitHub Actions job that prepares the environment before a coding agent starts work. Teams write one when the agent needs project dependencies installed before it can build and test. The job lives in a special GitHub Actions workflow file located at .github/workflows/copilot-setup-steps.yml, and its steps are executed in GitHub Actions before Copilot starts working. This customization applies only when Copilot runs on GitHub Actions and doesn't apply to sessions in a cloud sandbox.
Custom instructions
Custom instructions are standing guidance given to an AI assistant so you need not repeat the same context in every prompt. GitHub's docs say repository custom instructions files give Copilot additional context on how to build, test and validate changes. Repository-wide ones go in a copilot-instructions.md file in the .github directory. Copilot also reads agent instructions from files called AGENTS.md, CLAUDE.md, or GEMINI.md, which are currently not supported by all Copilot features.
Dev container
A dev container is a container configured as a project's development environment, so people and agents run the same tooling. For agent work it doubles as an isolation boundary. Claude Code's sandboxing page links to a page that compares the built-in sandbox with dev containers, containers and VMs, and names the dev container with its firewall script as one way to put a boundary around the whole process. Its permission modes page lists .devcontainer among protected directories.
Draft pull request
A draft pull request is a pull request marked as work in progress, opened for discussion before it is ready to merge. GitHub says drafts are useful when you want to share work-in-progress without formally requesting reviews. It also says draft pull requests cannot be merged, and code owners are not automatically requested to review them. Where a repository requires approvals, your approval of a Copilot pull request won't count toward the required number, so another reviewer must approve it before it can be merged.
Git worktree
A Git worktree is an extra working directory tied to one repository, so several branches can be checked out at once. Engineers reach for it when parallel agents must not overwrite each other's files. Git's manual says a repository can support multiple working trees. Claude Code describes one as a separate working directory with its own files and branch and starts an isolated session in one with the --worktree flag. The worktree guide covers the workflow.
Headless mode
Headless mode runs a coding agent without a person at the terminal, usually from a script or CI job. Claude Code's workflow docs say to run Claude non-interactively for CI, pre-commit hooks, or batch processing. In -p mode, user-invoked skills and custom commands work, while built-in commands that only run in the terminal interface, such as /login, aren't available. The headless agents guide goes further.
Hooks
A hook is a user-configured command or handler that an agent runs automatically when a named event occurs in a session. Hooks let a check such as a linter run at a fixed point instead of relying on the agent to remember it. In Claude Code, command hooks receive JSON data via stdin. Events such as PreToolUse and PostToolUse filter on tool name through a matcher, and the docs show a hook that runs a linting script only when Claude writes or edits a file.
Landlock
Landlock is a Linux kernel security feature that lets a process restrict its own access to resources such as files. It becomes relevant when comparing how agents are confined on Linux and on macOS. On Linux and WSL2, a masked credential file gives Claude Code's sandboxed commands a sentinel copy of the file, and the proxy substitutes the real value on outbound requests. In Codex, the default workspace-write policy still includes protected paths inside writable roots, and protection is recursive, so everything under them is read-only.
Merge conflict
A merge conflict is the state Git reaches when two branches change the same part of a file and the conflicting change has to be resolved before the merge can finish. GitHub counts conflicts among the blockers for Agent Merge, in which Copilot watches a pull request and works on the failed GitHub Actions workflows, merge conflicts, or both that you select. You turn it on with Watch and fix in the pull request's merge box.
Output style
An output style is a saved set of instructions that shapes how an agent responds throughout a session. You reach for one when you want shorter answers or more explanation without repeating the request in each prompt. Claude Code defines an output style as a set of instructions for role, tone, and response format. The same page says a style doesn't guarantee that something always happens or never happens.
Permission mode
A permission mode is a session setting that decides which agent actions need your approval and which run without asking. You choose one when deciding how much an agent may do on its own. Claude Code's permissions page says modes other than Manual change which actions ask for approval, and that in auto mode a classifier reviews actions instead of you. Its hooks reference reports the current mode as one of default, plan, acceptEdits, auto, dontAsk, or bypassPermissions.
Plan mode
Plan mode is a state in which a coding agent reads and proposes but edits nothing until you approve. It suits a risky change where you want to read the agent's intended edits first. In Claude Code, Claude reads files and proposes a plan but makes no edits until you approve, and the /plan command switches into plan mode ahead of a large change.
Plugin
A plugin is a bundle of agent extensions that installs and loads as one package. Engineers meet one when a teammate hands over a whole setup instead of loose files. Claude Code's docs say to use a plugin when you want several skills, subagents, hooks, or MCP servers packaged as one unit. At project scope a plugin is enabled for everyone in the repository, though each collaborator still installs it. Before installing, know that a plugin can run code with your user privileges.
Project rules
Project rules are instruction files stored with a repository that tell an agent how to work on that codebase. Teams commit them so every contributor's agent starts from the same conventions. In Cursor, project rules live in .cursor/rules as .mdc files and are version-controlled. Cursor also says a plain .md file there is ignored by the rules system because it has no frontmatter, and it points to AGENTS.md for anyone who prefers plain markdown.
Prompt injection
Prompt injection is an attack in which crafted input changes how a language model behaves. OWASP says prompt injection occurs when user prompts alter the LLM's behavior or output in unintended ways. Its page adds that indirect prompt injections occur when an LLM accepts input from external sources, such as websites or files. Codex's docs advise caution when enabling network access or web search, and warn that prompt injection can cause the agent to fetch and follow untrusted instructions. Claude Code's docs say sandbox restrictions still apply even if an injection bypasses Claude's decision-making.
Repository map
A repository map is a compact summary of a codebase's files and key symbols that a coding tool sends to a model as context. It helps most in large repos, where the model has to find the right files before editing. Aider uses a concise map of the git repository holding the most important classes and functions with their types and call signatures, and it sends a repo map with each change request. The map's size follows a token budget, and the --map-tokens switch defaults to 1k tokens.
Rewind
Rewind is a command that restores earlier code, the conversation, or both from a checkpoint saved during an agent session. You reach for it after an agent edit goes wrong and you want the files back without a Git reset. In Claude Code, /rewind rolls code and conversation back to a checkpoint, or summarizes part of the conversation. Checkpointing only tracks files edited within the current session, and a file that only Bash commands changed isn't tracked and stays as it is, so Git still matters.
Sandboxing
Sandboxing runs an agent's commands inside operating-system limits on files and network, so a bad command has a smaller reach. It becomes relevant once an agent may run shell commands without a person approving each one. Claude Code's sandboxing page says the sandbox isolates shell commands and their child processes, and its permissions page lists it as OS-level filesystem and network isolation for Bash commands. The sandboxing page also states that effective sandboxing requires both filesystem and network isolation.
Seatbelt
Seatbelt is the macOS sandbox facility that limits what a process can touch on the filesystem and network. Engineers run into the name when setting up coding agents that confine shell commands on a Mac. Claude Code describes its sandboxing as OS-level enforcement that restricts shell commands' filesystem and network access. On macOS, when a credential file is masked, sandboxed commands can't read the file at all, so tools that authenticate with it don't work inside the sandbox.
Secret scanning
Secret scanning is a check that finds credentials, such as API keys and passwords, committed to a repository. It matters with agents that write and commit files quickly, because a pasted key can reach history before anyone reads the diff. GitHub warns that committed hardcoded secrets become targets for unauthorized access. Its secret scanning automatically detects credential leaks so you can secure them before they're exploited.
Slash command
A slash command is an instruction typed after a slash in an agent session to control the session or start a workflow. In Claude Code, commands control Claude Code from inside a session, and you can type / to see the commands available to you. A command is only recognized at the start of your message, and text after its name becomes its arguments. The docs point to skills for readers who want to create their own commands.
Spec-driven development
Spec-driven development is a method that treats a written specification as the main artifact and the code as output generated from it. It shows up when agents build features from reviewed plans instead of one-off prompts. GitHub's Spec Kit methodology says the specification becomes the primary artifact, with code as its expression in a particular language and framework. The Spec Kit README tells users to define what to build and why, then turn the specification into a plan, tasks and working code.
Status check
A status check is a pass or fail result that an automated job reports against a commit or pull request, often used to gate merging. On GitHub, the Checks tab shows automated tests, builds, and other validations. With agent work, GitHub's Copilot docs say that by default GitHub Actions workflows will not run automatically when Copilot pushes changes, so confirm which checks actually ran before you read a green result.
Subagent
A subagent is a separate agent that a main session hands a task to, so the side work stays out of the main conversation. It pays off when a search or test run would flood the main thread with output. Claude Code's documentation calls subagents specialized AI assistants for specific types of tasks. Each runs in its own context window with specific tool access, and subagents work within a single session.
Suggested change
A suggested change is a reviewer's proposed edit, attached to a review comment, that the author can accept as a commit. It lets a reviewer hand an agent's pull request an exact fix in place of a description. GitHub says people can suggest changes, and you can apply these suggested changes directly if you have write access. Applying one or a batch creates a single commit on the compare branch.
Unified diff
A unified diff is a text patch format that lists changed lines together with a few unchanged lines around them, marking what was removed and what was added. Engineers meet it when an agent returns an edit as a patch to apply, or when they read a change in review. Pull request review also runs on diffs, and on GitHub the Files changed tab shows the diff that reviewers use to understand the proposed changes.