Vibe Coding vs Agentic Coding: What Actually Changes

What Vibe Coding Means

Karpathy coined the term in a post on X on 2 February 2025. He described a new kind of coding where you "fully give in to the vibes, embrace exponentials, and forget that the code even exists."

The rest of the post is the clearest definition. He spoke to Cursor Composer by voice and accepted every change without reading the diffs. He pasted error messages straight back in without comment. When the model could not fix a bug, he worked around it, or kept asking for changes at random until the bug disappeared. He admitted the code outgrew his understanding, and said the approach was fine for weekend projects you plan to throw away.

In short, vibe coding means steering by what you see on screen, not by reading the code.

What Agentic Coding Means

The vendors describe their coding agents in similar terms:

• Claude Code. Anthropic calls it "an agentic coding tool" that reads your codebase, edits files and runs commands, in the terminal, an IDE, a desktop app or the browser (overview). Its docs describe a repeating loop: gather context, take action, verify results (how Claude Code works).
• GitHub Copilot cloud agent. It can research a repository, plan changes and implement them in the background, editing files and running tests and linters in an ephemeral cloud development environment. Copilot handles the branch, commit messages and pushes, and can open one pull request per task (Copilot on GitHub.com).
• OpenAI Codex. OpenAI's docs list it as a CLI, an IDE extension and Codex Cloud. Each cloud task gets its own workspace, and you inspect the changed files, ask for follow-ups, then commit or open a pull request when you are ready (Codex Cloud).

A year later, in a post on 4 February 2026, Karpathy wrote that programming through LLM agents was becoming a normal way for professionals to work, but with closer supervision. The goal, he said, is to get the leverage of agents without letting software quality slip. His favourite name for it is agentic engineering: you mostly do not write the code yourself, you are "orchestrating agents who do and acting as oversight."

Agentic coding, then, is not vibe coding with a better model. It is delegation with a review step built in.

Vibe Coding vs Agentic Coding Side by Side

Sources: Karpathy's 2025 post, How Claude Code works, Claude Code on the web, GitHub's risks and mitigations for Copilot cloud agent, OpenAI's agent approvals and security, and our post on agentic pull requests.

The tool does not decide which column you are in. Lovable calls its Build mode (previously Agent mode) its autonomous execution mode, and Replit's assistant is called Replit Agent. You can vibe code with Claude Code by approving everything unread. The process around the tool decides.

What Actually Changes When You Move to Coding Agents

Going from prompting an app builder to running coding agents changes three things.

The code becomes yours to read. In a builder you judge the preview. With an agent you get a diff or a pull request, and that is the cheapest place to catch a problem. Reviewing the diff is rule 2 in our vibe coding best practices.

The agent can act, not only write. Locally, Claude Code can be allowed to run any command you can: build tools, git, package managers, scripts. Its checkpoints undo file edits, but actions on remote systems such as databases, APIs and deployments cannot be checkpointed (source). GitHub also warns that people can hide messages in issues or comments meant for its agent, a form of prompt injection (source). For the mistakes agents make in the code itself, see AI coding agents security risks.

Your main branch needs a gate. Lovable's GitHub sync edits one branch at a time, by default the repository's default branch (usually main). If branch protection rejects a push, Lovable puts the work on a lovable-sync branch for you to review and merge instead. Once agents are opening branches too, main is where you decide what ships.

How to Change Your Process

1. Protect main. Turn on branch protection and require status checks to pass, so nothing merges until your tests do. GitHub's docs say protected branches work on public repositories with GitHub Free, and on private ones with GitHub Pro, Team or Enterprise (checked on 24 October 2026). With a co-founder or contractor, also require an approving review.
2. One task, one branch, one pull request. Small pull requests are ones you will actually read.
3. Write down what done means. Claude Code reads CLAUDE.md at the start of every session, or an AGENTS.md in its place, and hooks can run shell commands before or after its actions. Our guide to making Claude Code or Cursor check its own code shows how to make it pass your typecheck and tests before stopping.
4. Set permissions before you need them. See the settings below.
5. Read the pull request yourself. Copy GitHub's rules for its cloud agent: the person who asked for the pull request cannot approve it, and by default workflows do not run on it until someone with write access approves them. OpenAI's docs put it plainly: "Treat Codex suggestions like any other PR". If an agent writes the code and a bot approves it, read when the AI approves its own pull request.

# BAD: skip permission prompts on your own machine, then push straight to main
claude --dangerously-skip-permissions "add Stripe billing"
git push origin main

# GOOD: plan first, work on a branch, run your own tests, open a draft pull request
git switch -c agent/stripe-billing
claude --permission-mode plan
npm test
gh pr create --draft --fill

The --dangerously-skip-permissions flag is the same as the bypassPermissions mode (CLI reference), which Anthropic says to use only in isolated environments such as containers or VMs. Since v2.1.283, Claude Code starts interactive terminal and VS Code sessions in Auto mode by default, where a classifier reviews most actions and blocks risky ones instead of asking you (source). Your own rules go in .claude/settings.json, as in this example adapted from the permissions docs:

{
  "permissions": {
    "allow": ["Bash(npm run *)", "Bash(git commit *)"],
    "deny": ["Bash(git push *)", "Read(./.env)"]
  }
}

Deny rules are checked first, and no allow rule can override one. But a Bash rule matches the usual form of a command, not every form, so the docs say it is not a security boundary; for hard limits they point to the sandbox. On your machine, Codex defaults to an OS-enforced sandbox with no network access and writes limited to the workspace, and asks before going beyond either (source).

Tools like VibeDoctor's Vibe Check (vibedoctor.io) automatically scan the repository your agent works in and your live site, and flag committed secrets, dependency CVEs and imports of packages that do not exist, with specific file paths and line numbers. Agents can call it over MCP from Claude Code, Cursor or Codex. Free to sign up.

When Vibe Coding Is Still Fine

Karpathy drew the line himself. Looking back in 2026, he wrote that with the models of early 2025, vibe coding was mostly used for fun side projects, demos and experiments. That is still a good use: a prototype to show a customer, or a tool only you will run.

The switch is due when real users, real data or payments arrive. From then on, someone has to read what ships, and that someone is you.

FAQ