What is the difference between vibe coding and agentic coding?
Published September 30, 2026 · by the AQ team
Vibe coding is building software by prompting an AI and accepting the generated code without reading it: you judge the work by running the app, not by reviewing the diff. Agentic coding is delegating a goal to an AI coding agent (Claude Code, Codex CLI, Cursor Agent, OpenCode) that plans, edits files, runs commands, and iterates on its own, while a person stays accountable for what ships. The difference is not the tool, since the same agent CLI can serve both modes; it is where human attention goes. Vibe coding removes review to maximize speed. Agentic coding moves the human out of the keystroke loop and into supervision: writing the goal, watching the run, and reviewing the result before it merges.
Who coined each term, and when
Vibe coding has a precise birthday. Andrej Karpathy coined it in a post on X on February 2, 2025: "There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists." The word spread far beyond programming circles; Collins English Dictionary named vibe coding its Word of the Year on November 6, 2025. Karpathy's original post was explicit that the mode was for "throwaway weekend projects", a caveat most later usage dropped.
Agentic coding has no single coiner; it emerged as vendor and research vocabulary in the same year. The earliest widely read use is Anthropic's launch of Claude Code as "an agentic coding tool" on February 24, 2025, three weeks after Karpathy's post. The contrast between the two terms was formalized in a May 26, 2025 arXiv paper by Sapkota, Roumeliotis, and Karkee ("Vibe Coding vs. Agentic Coding: Fundamentals and Practical Implications of Agentic AI"), which defines vibe coding as intuitive, conversational, human-in-the-loop work and agentic coding as goal-driven agents that plan, execute, test, and iterate with minimal human intervention. "Agentic" itself is the adjective the industry settled on for AI systems that take multi-step actions toward a goal rather than answering one prompt at a time.
The difference at a glance
As of September 2026, the practical differences look like this:
| Topic | Vibe coding | Agentic coding |
|---|---|---|
| Unit of work | A prompt and its reply, one exchange at a time | A delegated task: the agent plans, edits, runs tests, and retries until done or stuck |
| Who reads the code | Nobody; the app either works or you re-prompt | A person reviews the diff (and increasingly the session that produced it) before merge |
| Accountability | Explicitly surrendered; that is the point of the name | Kept by the human, who answers for the change like any other commit |
| What persists | A prototype, often regenerated rather than maintained | A repository with history, tests, and CI that the team maintains for years |
| How many at once | One conversation | Often several agents in parallel, one isolated checkout each |
| Where it runs | A browser app builder or one laptop terminal | Laptops, and increasingly cloud machines where long runs survive a closed lid |
| Failure mode | An app that works in the demo and cannot be safely changed | A plausible-looking diff that merges without being understood |
Who reviews the work
Review is the cleanest dividing line. In vibe coding, the generated code is treated as an implementation detail of the prompt: if the app behaves, the code is accepted, and if it misbehaves, the fix is another prompt, not a debugging session. That trade is honest for what Karpathy described, disposable projects where the cost of a wrong answer is near zero. In agentic coding the code is the deliverable. The agent may write every line, but a person reads the change before it lands, because the repository has users, teammates, and a future. The craft that has grown around agentic coding in 2025 and 2026 (instruction files, permission modes, session review, smaller scoped tasks) exists to make that review load bearable rather than to eliminate it.
What persists afterward
Vibe coding optimizes for the artifact you can see today: a working prototype, a landing page, an internal tool. Regenerating it from a better prompt is often cheaper than maintaining it, so version history, tests, and documentation matter little. Agentic coding assumes the opposite: the output joins a long-lived codebase, so the agent is judged by whether its change survives the test suite, the reviewer, and the next six months of edits. This is why agentic workflows lean on the repository's own guardrails (branches, CI, code review) instead of replacing them, and why teams describe the practice as a lifecycle rather than a trick; the agentic development lifecycle page covers that framing.
How many agents, and where they run
A vibe coding session is a dialogue, so there is exactly one of it. Agentic coding parallelizes: once the human's job is writing goals and reviewing results, nothing limits the number of concurrent agents to one, and developers commonly run several tasks at once in separate git worktrees so the checkouts never collide (see running multiple coding agents in parallel). Placement follows the same logic. A conversation happily lives in a browser tab; a forty-minute autonomous run does not want to die because a laptop lid closed, which is why agentic work is migrating to always-on machines. The agent itself is a coding harness, the loop of prompt, tools, and context wrapped around a model, and the harness is the same software in both modes; what changes is how much of the loop you delegate and how much infrastructure the delegation deserves.
The vocabulary keeps moving
Both terms are drifting toward the same middle. On October 7, 2025, Simon Willison proposed "vibe engineering" for the professional end of the spectrum: accelerating work with LLMs while "staying proudly and confidently accountable" for the result. In February 2026, Karpathy himself called vibe coding passé, saying programming via LLM agents had become a default professional workflow "except with more oversight and scrutiny", and offered "agentic engineering" as his preferred name for it: agentic because you are mostly orchestrating agents rather than typing code, engineering because there is expertise to doing it well. Whatever label wins, the direction is one-way: the industry is converging on delegation with oversight, and pure review-free vibe coding is retreating to the throwaway projects Karpathy originally scoped it to.
Where the team layer fits
Agentic coding at team scale has a visibility problem that vibe coding never had: a July 2026 LeadDev analysis of 25,264 agent-generated pull requests across 2,361 popular GitHub repositories found that in 79 percent of agentic PRs the same developer both reviewed and modified the agent's contribution, and only about one in eight workflows involved multiple humans. Delegation with oversight only works when someone other than the delegator can actually see the work, which is why teams that run many agents end up organizing the practice deliberately; the software factory page describes that end state.
Where AQ fits
AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud. In this page's terms, AQ is infrastructure for the agentic side: agents run as real CLIs (Claude Code, Codex, Cursor Agent, Kimi, Grok, or plain shells) in persistent tmux sessions on your team's VM, so a long run survives a closed laptop and resumes from any device. Each workspace is one isolated git worktree, which is what makes several parallel agents practical, and teammates open the same workspace and watch the same live session, so the review that separates agentic coding from vibe coding can be done by a second person, not just the delegator. Everyone signs in with their own Claude or OpenAI account (AQ never marks up usage on your own subscriptions), agents commit, push, and open PRs with per-user GitHub auth, and a live dev-server preview per workspace keeps the vibe coder's feedback loop, seeing the app run, without giving up the diff. Pricing is two plans: Free is a personal sandbox for one person (AQ creates a private machine in an isolated network, nothing to install, no time limit), and Team is $50 per user per month in early access (standard $200, billed monthly), covering VMs you connect from your own cloud or a dedicated always-on AQ-managed VM in its own isolated network, rate locked for your first 12 months.
Frequently asked questions
Is vibe coding the same as using Claude Code or Codex?
No. Claude Code and Codex CLI are agent harnesses, and the same harness serves both modes. If you accept their output without reading it, you are vibe coding; if you delegate a goal and review the diff before it merges, you are doing agentic coding. The term describes the human's posture toward the code, not the tool.
Who coined vibe coding and agentic coding?
Andrej Karpathy coined vibe coding in an X post on February 2, 2025. Agentic coding has no single coiner: Anthropic launched Claude Code as "an agentic coding tool" on February 24, 2025, and a May 2025 arXiv paper by Sapkota, Roumeliotis, and Karkee formalized the vibe coding versus agentic coding contrast. The adjective "agentic" was already standard AI vocabulary for systems that take multi-step autonomous actions.
Is vibe coding bad?
Not for what it was named for. Karpathy scoped it to throwaway weekend projects, and for prototypes, demos, and personal tools it is the fastest way to get something running. It becomes a problem when review-free code lands in software other people depend on: unread code can carry security holes, licensing surprises, and designs nobody can maintain. The rule of thumb: vibe code anything you could regenerate from scratch, review everything you intend to keep.
What did Karpathy mean by vibe coding being passé?
In February 2026, a year after coining the term, Karpathy said models had improved enough that programming via LLM agents was becoming the default professional workflow, "except with more oversight and scrutiny", and said his preferred name for that discipline was "agentic engineering". He was not disowning the original idea; he was noting that the interesting frontier had moved from review-free generation to orchestrating agents while keeping software quality.
Can a team do agentic coding without everyone watching each other's agents?
It happens by default, and it shows: a July 2026 LeadDev analysis of 25,264 agent-generated pull requests found that in 79 percent of agentic PRs the same developer both reviewed and modified the agent's contribution. That collapses the oversight that distinguishes agentic coding from vibe coding into one person. Teams fix it with shared visibility: agent sessions that a second person can open, watch, and review, rather than terminals locked to one laptop.