Vibe coding tools: the honest list, sorted by what they actually do

What vibe coding actually means

Andrej Karpathy posted the phrase on 2 February 2025, describing a way of working where you "fully give in to the vibes, embrace exponentials, and forget that the code even exists." His own summary of the loop was: "I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works."

Collins Dictionary named it Word of the Year on 6 November 2025, glossing it as "programming by vibes, not variables." That is the popular definition, and it is worth being precise about what it excludes.

Karpathy’s own framing included the caveat everyone quotes selectively: it works for "throwaway weekend projects." The interesting question for 2026 is what happens when people use it for things that are not throwaway, which is most of what actually shipped.

Lane one: app builders

You describe an app in a chat box and get a deployed, running application with a database and auth wired in. No local environment, often no git until you export. This is the purest expression of the idea and by far the largest lane by user count.

The other structural issue is the exit. These tools are excellent at zero to one and progressively worse from one onward, because the assumption that you are not reading the code stops holding the moment you need a change the model cannot make. Check what the export looks like before you commit a real project, not after.

Lane two: AI editors

An editor with a model wired into it. You still see files, still have a terminal, still commit yourself, but completion, chat, and an in-editor agent do most of the typing. Cursor defined this lane. Devin Desktop is what Windsurf became after Cognition acquired it.

This is the lane where the word "vibe" is least accurate, because the code is right there on screen. That is a feature. The editor lane is the smallest jump from how you already work, and the easiest to dial the autonomy up and down on a per-task basis.

The cost model differs from lane one in a way that trips people up: editors charge per seat with request or usage limits behind them, so the marginal cost of a long debugging session is bounded by your plan rather than by a draining credit balance. See AI coding pricing compared for the seat-versus-token arithmetic.

Lane three: terminal agents

Claude Code and Codex run in a terminal inside a real repository. They read the tree, run your tests, edit files, and hand you a diff. There is no editor and no GUI in the box.

These are the most capable tools in the list on hard, existing codebases, and the least forgiving of the vibe-coding behaviour, because the output is a commit against code other people depend on. They are also the only lane where the agent can genuinely run your test suite and iterate on real failures rather than on a sandbox approximation of your app. Which model you point one at moves the result as much as which tool you picked, and our coding-model leaderboard ranks the models these agents run on by measured pass rate and cost per task.

Lane four: orchestration

One agent in one terminal is a bottleneck once you trust it. The fourth lane runs several at once, each isolated in its own git worktree and branch, with a review surface in front of them: Continuum, Conductor, and T3 Code.

This lane barely existed when Karpathy coined the phrase, and its existence is the clearest evidence of where the practice went. Nobody builds a fleet-management layer for a workflow where you are not reading the output. The whole point of these tools is the review, the isolation, and the ledger.

Where vibe coding breaks

Four failure modes, in rough order of how often they bite.

1. The debugging cliff. The model writes something that mostly works, then cannot fix the part that does not, because it cannot see the failure the way you can. On a credit-metered builder, each attempt costs money and the price of the fix can exceed the price of the build.
2. Security you did not read. Unreviewed code ships whatever the model assumed about auth, input validation, and secrets. This is the single most cited real-world consequence and the reason agent security stopped being a niche topic.
3. Architecture drift. Twenty accepted suggestions produce twenty locally reasonable decisions and one incoherent codebase. Nobody notices until a change has to cross three of them.
4. Nobody understands the system. The most expensive one, because it surfaces months later when the person who vibed it is asked a question about it.

The common thread is that all four are review failures rather than model failures. That is why the practice professionalised the way it did: not by generating less, but by putting a gate between generation and merge. Reviewing AI-generated code is the discipline; agentic coding is what the workflow is called once you have it.

Questions people ask

Keep reading

Continue through Agentic coding with Remote coding: the four patterns, and which one survives a dropped connection.