Agentic Coding: Beyond Vibe Coding

Agentic coding means an AI agent gets a task and works through it in a loop. It reads the existing code, plans the steps, changes files, runs commands and tests, reads the result and fixes what's wrong until the task is done. You set the goal and the limits and review the result. Unlike vibe coding, the code gets read, tested and reviewed. Unlike autocomplete in your editor, the AI doesn't just write the next line, it works through a whole task with tools.

What agentic AI is in general is in Agentic AI Explained: What It Is and Isn't. This post is about programming, and further down about how I work this way.

I collect the other building blocks and guides under Understanding and Using Claude Right.

What agentic coding is: the loop of plan, act, verify

Anthropic describes how Claude Code works as an "agentic loop" with three phases: gather context, take action, verify results. These phases blend together, the loop runs until the task is complete, and you can step in at any point.

Tools make this possible: according to the docs, without tools Claude can only respond with text; with them it reads code, edits files and runs commands. Given "fix the failing tests", the agent runs the test suite, reads the error output, finds and edits the relevant files and runs the tests again.

The developer Simon Willison calls code execution the defining capability that makes this way of working possible. A coding agent checks whether the code runs, not just whether it looks plausible.

Agentic engineering: where the term comes from

You'll also read "agentic engineering". Who used it first, I can't document, but two sources are clear. Andrej Karpathy, who had already coined "vibe coding" in February 2025, wrote a retrospective on X on February 4, 2026, for the first anniversary of vibe coding. Programming via agents, he wrote, is becoming a default workflow for professionals, with more oversight and no compromise on quality. Many people had tried to find a better name to set this apart from vibe coding; his current favourite was "agentic engineering". So he doesn't present the term as his own invention. His reasoning: "agentic" because you almost never write the code directly any more, you orchestrate agents and oversee them. "Engineering" because it takes expertise you can learn and get better at.

Simon Willison has been collecting patterns for it since February 23, 2026, in his guide "Agentic Engineering Patterns". He defines agentic engineering as developing software with the assistance of coding agents, meaning agents that can both write and execute code. His examples are Claude Code, OpenAI Codex and Gemini CLI.

I use agentic coding and agentic engineering interchangeably here. To me, the word engineering stresses the discipline around it: the brief, tests, review, responsibility.

Agentic coding vs vibe coding vs autocomplete

The difference is how much the AI does on its own and who checks the result.

GitHub describes this range for Copilot itself: from suggestions in the editor to an autonomous agent that edits files, runs commands and validates its work.

Vibe coding isn't bad, it has a different purpose. Willison reserves the term for unreviewed, prototype-quality code. When that holds up is covered in What Is Vibe Coding? A Plain Guide. Agentic coding is for anything other people will use.

Tools for agentic coding (as of October 30, 2026)

• Claude Code from Anthropic runs in the terminal, the desktop app, IDEs, on the web and in CI pipelines, according to the docs with the same loop everywhere. The way in is Claude Code: A Guide for Beginners.
• Codex from OpenAI comes as a CLI, an IDE extension, in the ChatGPT app and as Codex Cloud. Locally, according to OpenAI, it works in a sandbox, typically limited to the current workspace, with network access off by default.
• Gemini CLI from Google is an open-source terminal agent under the Apache 2.0 licence. Since June 18, 2026, personal accounts and Google AI plans use Antigravity CLI instead; Gemini CLI now targets enterprise Gemini Code Assist customers.
• GitHub Copilot has, besides completions, an agent mode in the IDE and a cloud agent. According to GitHub, the cloud agent creates branches, writes code and opens pull requests in the background.

How Claude Code and Codex differ day to day is in Claude Code, Codex or Copilot: Which One?. All alternatives with dates and sources are in Claude Code Alternatives in 2026.

What a good agentic coding setup needs

To me, six building blocks belong in it:

1. A clear brief. Anthropic recommends naming the goal and how to check it: instead of "fix the login bug", say "write a failing test that reproduces it, then fix it". For bigger tasks, put in writing beforehand what should be built and how you'll know it's done. How to do that as a method, with spec, plan, tasks and a check, is in Spec-Driven Development With AI Agents.
2. Context files. What the agent can't read from the code goes into a file it loads at every start: CLAUDE.md for Claude Code, AGENTS.md for Codex. According to Anthropic, it should only hold what Claude couldn't work out on its own. What goes where is in CLAUDE.md vs AGENTS.md: What Survives.
3. Permissions and sandbox. An agent should only be allowed what the task needs. Rules that must never be broken belong, according to Anthropic, in deny rules, hooks or the sandbox, not in the instructions. Deny rules hold in every permission mode. How that works in Claude Code is in Claude Code Security: Permissions and Sandbox and Claude Code Hooks: Rules That Always Apply.
4. Tests or another check. Anthropic writes that Claude stops when the work looks done. If it has nothing that tells it whether the work actually is done, you become the test suite. A test, a build or a linter that comes back pass or fail closes the loop.
5. A separate review. In Claude Code, /code-review reviews the changes in a fresh subagent that never saw your conversation. According to Anthropic, that's why it isn't swayed by the reasoning that produced the code. More in Claude Code Review: Cost, Setup, Limits and Claude Code Subagents: Best Practices.
6. Git. Claude Code can undo file edits with checkpoints, but nothing that happened outside, such as in a database or a deployment. Git records every state. Branches keep experiments apart from the finished state. If that's new to you, read What Is Git? Git and GitHub for Beginners.

How I use agentic coding

My AI employees work with Claude Code, each in their own folder with personnel file, working instructions, skills, knowledge and job folders. Before anything goes live, a separate reviewer checks it, and whoever created something never reviews it themselves. How such a folder is built is in Creating an AI Employee: How to Start; how the employees run on a server is in Claude Code on a Server: tmux, Remote Control. What a review against my brief looks like is shown on the page about Kai.

The risks of agentic coding

• "Done" isn't done. According to Anthropic, Claude sometimes reports finished work that isn't quite finished, usually because it couldn't check it.
• Prompt injection. Instructions can be hidden in untrusted content such as web pages. Anthropic advises not piping it directly to Claude and using virtual machines for scripts that call external services.
• Irreversible actions. What the agent triggers in databases, through APIs or in deployments isn't covered by any checkpoint. You control that only through permissions and rules.
• Automatic approvals can be wrong. In Claude Code's auto mode a classifier decides instead of you. According to Anthropic, that's a judgment call, not a hard rule, and can be wrong.
• Responsibility. According to Anthropic, you're responsible for reviewing proposed code and commands before approval. OpenAI advises treating Codex suggestions like any other pull request.

Frequently asked questions

What is agentic coding?

Software development with an AI agent that reads code, edits files, runs commands and tests and iterates on its own until a task is done, while you set the brief and review the result.

What's the difference between agentic coding and vibe coding?

With vibe coding you accept the AI's code without reading it. With agentic coding an agent also works largely on its own, but with tests, rules and a review by you or a second role. Vibe coding suits prototypes, agentic coding suits software other people use.

What does agentic engineering mean?

Agentic engineering is software development with coding agents treated as a serious discipline. Andrej Karpathy named it in February 2026 as his favourite term for the professional way of working after vibe coding, and Simon Willison defines it in his guide "Agentic Engineering Patterns".

Which tools are there for agentic coding?

Well-known tools are Claude Code from Anthropic, Codex from OpenAI, Gemini CLI and Antigravity CLI from Google, and the agent mode and cloud agent of GitHub Copilot.

Do I need programming skills for agentic coding?

Not necessarily to start. But for it to be more than vibe coding, someone has to write a clean brief, set the checks and judge the result.

What to do next

Start with a small, real task in a folder under Git. Write into the brief how the agent can tell it's done, and have the result reviewed by someone who didn't write it. If you haven't set up Claude Code yet, start with Install Claude Code: Windows, Mac, Linux. When your briefs get bigger, Spec-Driven Development With AI Agents will help. How others build and review with agents is what we discuss in my community AI Employee Framework.