What is a coding agent? Claude, Codex, Cursor, Grok

Nov 5, 2026

A coding agent uses a model and tools to read a repo, edit files and run commands in a loop. How Claude Code, Codex, Cursor and Grok Build each do that job.

A coding agent is the software around a model that can work in a repository: read files, edit them, run commands, and keep going until the task is done or you stop it. The model picks the next step. The agent touches the project, asks you before a risky command, and shows you the diff.

The big players in AI coding is the map of who builds the models and the products. Agentic coding is the loop those products run. This entry is the short definition of the software itself.

How a coding agent works

You describe an outcome. The agent looks at the project, takes a step, reads the result, and decides again. You can interrupt, redirect, or reject a command, depending on the permission mode.

Cursor’s agent page splits the setup into instructions, tools and the model you picked. Codex’s guide splits it into a model, tools, and the guardrails and review you keep. Claude Code’s glossary calls the software around the model the harness: file access, the shell, permission checks, memory, and the loop. Those are three ways of listing the same parts.

A chatbot that only returns text is not this, even when the text is code you paste in yourself. Vibe coding is the other neighboring term: a tool that scaffolds an app while you stay out of the repository.

Coding agents in each tool

	Claude Code	Codex	Cursor	Grok Build
What the docs call it	An agentic coding tool. The harness around Claude	An agent for code, a codebase, or developer tools	Agent, opened with Cmd+I	A coding agent
Where you run it	Terminal, IDE, desktop app, and the browser	App, CLI, IDE extension, and Codex cloud	Editor side pane, Agents Window, CLI	Terminal UI, `grok -p`, or ACP in another app
What it can do	Read files, run commands, edit	Inspect code, edit, run commands in a sandbox	Search, edit, terminal, browser, questions	The overview does not print a tool list
A limit the docs state	You can interrupt the loop	Local commands stay in the sandbox until you approve a step outside it	No limit on the number of tool calls in one task	Not documented on the overview

Claude Code

Claude Code is agentic because it has tools that let it act, not only advise. The loop the glossary describes is gather context, take action, verify, and repeat. You can interrupt that loop and redirect it.

The same engine runs in the CLI, VS Code, JetBrains, the desktop app, and claude.ai. A session on your machine reads your local CLAUDE.md, settings and skills. A cloud session starts from a fresh clone and does not read ~/.claude/ on your machine. See the Claude Code glossary.

OpenAI Codex

Codex is the product you use when the work is code. The prompting guide says a useful request names the behavior, points at the code or the steps that reproduce a bug, keeps the constraints, and says how to verify the change.

Local commands run inside a sandbox that limits file and network access. If a task needs to cross that boundary, Codex follows your approval policy before it continues. Cloud tasks run in a cloud environment, which the agents guide describes as inspecting code, making changes, and running commands or tests. See Working with agents and Prompting.

Cursor

Agent is the assistant that can complete coding tasks, run terminal commands and edit code. Cmd+I opens it in the side pane. Tools include search, web, rules, reading files, edits, the terminal, the browser, image generation and clarifying questions. The agent page says there is no limit on how many tool calls one task can make.

While Agent is working, Enter queues your next message. Cmd+Enter sends it immediately. A follow-up can also steer the current turn at the next tool call, without cutting off the action in progress. See Cursor Agent.

Grok Build

The overview calls Grok Build a coding agent you use from an interactive terminal UI, from a script with grok -p, or from another app through ACP. grok inspect lists the instructions, skills, plugins, hooks and MCP servers it found in the current directory.

The overview does not list file-edit and shell tools one by one. The feature pages it links cover skills, plugins and modes. See Grok Build.

Details checked against each vendor’s documentation in November 2026.

A real example

One request, the kind Codex’s prompting page asks for: the behavior, the place, the boundary, and the check.

The RSS feed at /rss.xml is missing posts that have an updated: field and no other change.

Look in src/pages/rss.xml.ts and src/posts/. Include a post when its date is in the past, whether or not updated: is set. Do not change the feed item limit.

Run npm run build and say whether the build finished.

Claude Code, Codex, Cursor and Grok Build can all take that prompt in the site’s repository. What you do next differs by product. In Claude Code you watch the tool calls and the permission prompts. In Codex the commands stay sandboxed until a step needs approval. In Cursor you can queue a second instruction while the first is running. In Grok Build you can send the same words to grok -p if you want the run in a script instead of the terminal UI.

Plan mode is the better first step when you do not yet know which file builds the feed.

When to use a coding agent

Use one when the next step depends on what is in the repo. A bug you can reproduce, a page that should match a pattern already in the code, a rename across files. Codex’s guide says to delegate when you can describe what done means, the agent can reach the files and commands it needs, and you can judge the result.

Skip it when you want a single explanation and you will apply the change yourself. Skip it when the job is “make me an app” and you do not have a repository yet. That is the case the vibe coding entry points at a different kind of tool for.

Pitfalls

You still review the diff. Codex’s guide says fluent writing is not proof the task succeeded, and that changing an instruction does not undo a command that already ran. If the agent reports that it updated something, look at the file.

A cloud session is not your laptop. Claude Code cloud sessions do not read ~/.claude/ on your machine, so a personal instruction that your local sessions follow can be missing there. Codex cloud is a separate environment from the sandbox on your computer. Say which one you are starting.

Cursor will keep calling tools until the task ends or you stop it. The page states no cap on that number. A vague prompt can spend a long turn searching. Name the file or the check.

Agents can take longer than a one-shot answer, and they can use more of your plan. The Codex guide says to check the pricing page before a large or repeating job. This entry does not copy those prices.

The big players in AI coding separates the model, the agent software and the interface, and names the current products.
How to use Grok Build is the long walkthrough of one of the four.
Keep a timeline of what your coding agents did today records the work after the session. Work Tracebook is the app that draws that timeline.
Help your coding agent find the CLI tools you already installed is how an agent discovers the commands on the machine, instead of guessing a package name.

Agentic coding: Agentic coding means directing an agent that reads your repo, edits files and runs commands. How that loop works in Claude Code, Codex, Cursor and Grok Build.
Permission modes: Permission modes decide what a coding agent can do without asking. The modes in Claude Code, Codex, Cursor and Grok Build, and what YOLO mode really skips.
Plan mode: Plan mode makes a coding agent research and propose a plan before it edits code. How it works in Claude Code, Codex, Cursor and Grok Build, and when to use it.
Vibe coding: Vibe coding is describing an app to a tool that scaffolds it while you stay out of the repo. How that differs from Claude Code, Codex, Cursor and Grok Build.

All AI glossary entries →

Follow @flaviocopes

Want me to talk about your product? You can sponsor this site.