Skip to main content

Agentic Coding Glossary: 40 Terms Explained

By the OnlineAppCreator editorial team4 min read

Short answer

Agentic coding is software development in which an AI agent plans, writes, runs and corrects code on its own within limits set by a developer. This glossary defines 40 terms that come up when teams adopt tools such as Claude Code, OpenAI Codex, Cursor and GitHub Copilot, grouped into core concepts, context, tools and control, workflows, and quality and security.

On this page
  1. Core concepts
  2. Context
  3. Tools and control
  4. Workflows
  5. Quality and security
  6. Using the glossary in your team

Teams adopting coding agents meet a new vocabulary within days: context windows, MCP, subagents, worktrees, approval modes. Shared definitions make policies, training and tool comparisons much easier. This glossary explains 40 terms in plain language, grouped by theme. Definitions are tool-neutral; where a term is specific to one product, we say so.

Core concepts

Agentic coding: development in which an AI agent carries out multi-step coding tasks, not just suggestions. Coding agent: a program that combines a language model with tools to read, edit and run code. Agent loop: the cycle of planning, acting, observing results and adjusting until the task is done. Large language model (LLM): the model that generates the agent's reasoning and code. Autocomplete: inline suggestions while typing, the pre-agent form of AI coding. Vibe coding: building software mainly through prompts while paying little attention to the code itself, fine for prototypes and risky for production. AI app builder: a hosted tool such as Lovable, Base44, Bolt or v0 that generates complete web apps from chat. Terminal agent: an agent that runs in a command-line shell, such as Claude Code or the Codex CLI. Agent mode: an editor feature that lets the built-in assistant act agentically, as in GitHub Copilot or Cursor.

Context

Context window: the amount of text a model can consider at once, measured in tokens. Token: a chunk of text, roughly part of a word, used to measure input, output and cost. Context file: a file in the repository that tells agents how the project works. CLAUDE.md: the context file Claude Code reads. AGENTS.md: an open context file convention read by Codex and several other tools; our guide to writing CLAUDE.md and AGENTS.md shows good structure. Rules: tool-specific instruction files, such as Cursor project rules or Copilot instructions. Compaction: summarizing earlier conversation to free space in the context window. Retrieval: searching code or documents to bring relevant parts into context. Brief: a written task description with goal, constraints and acceptance criteria.

Tools and control

Tool call: an action the model asks the agent to perform, such as reading a file or running a command. Model Context Protocol (MCP): an open standard for connecting agents to external tools and data; see MCP servers explained. MCP server: an adapter that exposes a tool or data source through MCP. Permission rules: settings that allow, ask about or deny specific agent actions. Approval mode: how much the agent may do without confirmation, from read-only to fully autonomous. Sandbox: an isolated environment that limits what agent commands can affect. Hook: a script that runs automatically at defined points in the agent's work, for example before a file edit; a Claude Code feature. Slash command: a reusable shortcut prompt invoked with a slash. Skill: a packaged set of instructions and resources an agent can load for a specific kind of task. Headless mode: running an agent non-interactively, for example in CI.

Workflows

Subagent: a helper agent with its own context that handles part of a task and reports back. Background agent: an agent that works asynchronously in a cloud environment and returns a pull request; see background and cloud coding agents. Git worktree: an additional working directory of the same repository, used to run agents in parallel on separate branches. Multi-agent workflow: several agents working in parallel or in roles such as writer and reviewer. Plan mode: a mode in which the agent proposes a plan for approval before editing anything. Test-first prompting: asking the agent to write failing tests before implementing a change. Human in the loop: a workflow in which people approve key steps such as plans and merges.

Quality and security

Hallucination: confident output that is wrong, such as an invented API or package. Prompt injection: malicious instructions hidden in content the agent reads, aiming to change its behavior. Slopsquatting: registering package names that models tend to invent, so agents install malware. Secret scanning: automated detection of credentials in code and commits. Guardrails: technical and procedural limits that keep agents within safe boundaries.

Using the glossary in your team

A shared vocabulary is the cheapest step in any rollout. A practical approach looks like this:

  1. Pick the 15 terms most relevant to your tools and add them to your internal AI usage policy.
  2. Reference terms such as approval mode, sandbox and MCP server explicitly in your security rules.
  3. Use the workflow terms when writing tickets for agents, so everyone means the same thing by plan mode or background task.
  4. Revisit the list twice a year, since the field adds new concepts quickly.

Teams that want to go from vocabulary to practice can book the hands-on agentic coding workshops from neooutput, available remote or on-site in English or German. At OnlineAppCreator, where AI Code Engineers take products from idea to operating business, these terms are part of everyday work.

Frequently asked questions

What is the difference between agentic coding and vibe coding?

Agentic coding describes the technique of letting an agent plan and execute coding tasks. Vibe coding describes a style of using it with little attention to the resulting code, which is acceptable for prototypes but not for production software.

Is AGENTS.md the same as CLAUDE.md?

They serve the same purpose. AGENTS.md is an open convention read by several tools, while CLAUDE.md is the file Claude Code reads. Many teams keep one and reference it from the other.

What does MCP stand for?

Model Context Protocol, an open standard that lets AI agents connect to external tools and data sources through MCP servers.

Want this done right, the first time?

OnlineAppCreator takes products from idea to operation. Tell us where you stand and get a concrete plan in a free 15-minute call.