Vibe, agentic or assisted: what’s the difference?

Agentic coding means an AI agent does the programming work in a loop: it reads your code, plans a change, edits files, runs commands and tests, checks the results and tries again until the task is done. AI-assisted coding means the AI suggests code while you write and approve every change. Vibe coding means you build by describing what you want and accepting the result without reading the code.

The first two terms describe how much the tool does. Vibe coding describes how much you check. That is why they are not three steps on one ladder, and why the same agent can be used for careful agentic coding or for pure vibe coding.

01What is agentic coding?

Agentic coding is software development where an AI agent works toward a goal on its own, using tools, and checks its own results before it stops.

The key word is loop. Anthropic’s documentation describes Claude Code’s agentic loop as three phases, gather context, take action and verify results, repeated until the task is complete. It adds that tools are what make Claude Code agentic: without them, Claude can only reply with text.

Simon Willison, a developer who writes widely on AI tools, put the general idea in one line in September 2025: “An LLM agent runs tools in a loop to achieve a goal.”

A concrete example: you type “fix the failing tests.” An agent runs the test suite, reads the errors, finds the files involved, edits them and runs the tests again, repeating until they pass. You review the result, not every keystroke.

Claude Code, OpenAI Codex, Cursor’s agent, GitHub Copilot’s agent mode and Google Antigravity all work this way. The term is now mainstream: Apple headlined its February 2026 release “Xcode 26.3 unlocks the power of agentic coding”, with Anthropic’s Claude Agent and OpenAI’s Codex built in. Our guide to the best AI coding agents of 2026 compares the tools.

02What is AI-assisted coding?

AI-assisted coding is programming where you stay the author and an AI helps, one suggestion at a time.

GitHub Copilot made this mainstream. Its June 2021 launch post said it draws context from the code you are working on, “suggesting whole lines or entire functions”. The phrase itself is older: on Hacker News it appears as early as July 2018, in a story about Microsoft’s Visual Studio IntelliCode.

One caution: people also use “AI-assisted” as an umbrella for every kind of AI help, agents included. Willison’s March 2025 essay, “Not all AI-assisted programming is vibe coding”, uses it that way. In this guide it means the narrow sense: suggestions you accept one by one.

03What is vibe coding?

Vibe coding is building software by describing what you want and accepting the AI’s code without reading it, judging the result only by running it.

Andrej Karpathy, a founding member of OpenAI, coined the term in a post on X at 23:17 UTC on February 2, 2025. We decoded that time from the post’s ID, since X IDs encode when a post was created.

He described accepting every change, “I ‘Accept All’ always, I don’t read the diffs anymore”, pasting errors back without comment, and letting the code grow beyond what he understood. He said it was not too bad for throwaway weekend projects.

One detail is often missed. His example tool was Cursor Composer with Claude Sonnet, which proposed whole diffs for him to accept, not line-by-line autocomplete. Vibe coding was born with agent-style tools, as a way of using them without review.

Willison then drew the line that most careful writers use. Vibe coding is “building software with an LLM without reviewing the code it writes.” If you reviewed it, tested it and can explain how it works, he argues, that is simply software development. Collins Dictionary made “vibe coding” its word of the year in November 2025, with a broader definition: using AI, prompted by natural language, to write code.

For the full story, risks included, see what vibe coding is and where it breaks.

04Is agentic coding the same as vibe coding?

No. Agentic coding describes what the tool does; vibe coding describes what you skip, which is reading the code. Put the two dials side by side and each term finds its place:

Most arguments about these terms mix up the two dials. An engineer running Claude Code with tests and a careful review is doing agentic coding. A first-time builder accepting everything an app builder produces is vibe coding. Both are using agents.

Karpathy made the same split a year after his first post. On February 4, 2026, he wrote that programming through agents was becoming a default for professionals, with more oversight and scrutiny. His favorite name for that is agentic engineering: “agentic” because “you are orchestrating agents who do and acting as oversight”, and “engineering” because it takes skill that you can learn.

05Where the terms came from

1. June 2021GitHub introduces Copilot, “your AI pair programmer”, which suggests whole lines and functions
2. June 2024Anthropic’s Claude 3.5 Sonnet launch reports an “internal agentic coding evaluation”; the phrase reaches Hacker News the same week
3. February 2, 2025Karpathy’s post coins “vibe coding”, and it reaches Hacker News about two hours later
4. February 24, 2025Anthropic introduces Claude Code as “a command line tool for agentic coding”
5. March 2025Simon Willison argues that not all AI-assisted programming is vibe coding
6. November 2025Collins names “vibe coding” its word of the year
7. February 2026Apple announces agentic coding in Xcode 26.3, and Karpathy proposes “agentic engineering”

To see how fast each term spread, we counted Hacker News stories and comments that use each phrase, through the public Hacker News search API run by Algolia, on September 24, 2026. Hacker News is one developer community, not the whole internet, and the counts include spelling variants such as “vibe-coding”. Read them as a trend, not a census.

Three things stand out. “Vibe coding” did not appear once before Karpathy’s post, and the first story linking to it came 2 hours and 10 minutes after he published. It is still the most discussed term. But “agentic coding” is growing faster: its monthly mentions rose 3.8 times from 2025 to 2026. And “agentic engineering” was almost unused before 2026, with 22 mentions in all of 2025.

06AI-assisted vs agentic vs vibe coding at a glance

07When should you use each?

Match the method to the cost of a mistake.

• Vibe code when a bug costs little: prototypes, demos, one-off scripts and tools only you use.

• Use agentic coding with review for anything with users, data or money. Write a short spec first, let tests define what done means, and review the diff by risk.

• Use AI-assisted coding when you are learning, or making small, precise edits where you want to type each line yourself.

Most real work mixes them. The AI pair programming workflow shows a loop that keeps an agent’s speed with a human’s judgment. To move a vibe coding session toward agentic coding with review, start with a prompt like this:

Before you change anything, write a short plan: the files you will touch, the tests you will add, and how I can check the result myself.
Wait for my approval.
Then work one step at a time. After each step, run the tests and show me the output.
Stop and ask me if a step would change login, payments, stored data or any file I did not mention.

08FAQ

Read next: what vibe coding is and where it breaks, or compare the best AI coding agents of 2026.