Code · guides

Vibe coding, AI agents and coding assistants

By Alberto Gulotta · Updated 9 September 2026 · 5 min read

This is the entrance to the code tower. Vibe coding — describing what you want and letting a model write it — assistants that sit inside an editor, agents that run for an hour on their own, and the tools that let people build without writing code at all.

The tower serves two readers who rarely appear on the same page. One is a developer deciding which assistant to put in their editor and how much to trust it. The other has never programmed, has an idea, and has just discovered that describing it to a model produces something that runs. Both are legitimate, the second is growing far faster, and pretending the second does not exist is how most technical writing on this subject fails.

What both need turns out to be the same thing, and it is what this building is planned around: knowing what to check. Generated code runs, looks finished and is frequently subtly wrong — a missing case, a secret left in a file, a dependency that nobody chose, an interface left open to the internet. None of that announces itself, and none of it is caught by the code working.

Agents are the newest part of this and the one where the gap between demonstration and daily use is widest. Anything written here about them will describe what they do, what they cost to run, and where the supervision has to sit, rather than repeating a launch video.

Building without code deserves its own place because for a great many problems the honest answer is that no code should be written at all — a form, a spreadsheet and an automation between two services solves it in an afternoon and can be maintained by somebody who is not a programmer.

Three levels are planned, and you are on the first. One guide exists so far, on what to do when Gemini CLI stops working, and it is linked below. The rest of the map is written out here as a plan: as each floor arrives it will cover one subject completely and hold the shorter guides underneath it, and nothing will be more than three clicks from the front door.

One observation deserves stating at the door, because it is the point on which this whole tower turns. The barrier to building software used to be writing it, and that barrier has genuinely fallen — somebody with no training can now describe an idea and watch something appear that runs. What has not fallen, at all, is everything that comes after: knowing whether it is correct, whether it is safe to expose to strangers, what it will do with data belonging to other people, and how it will be maintained in a year. Those were always the expensive parts of the job, and they are the parts the current generation of tools helps with least. That asymmetry is why checking is the largest part of the plan for this tower, and why it will be written to be read by somebody who has never programmed rather than by somebody who has.

Not covered here. It will not tell a beginner that generated code is production-ready. It frequently runs and it frequently is not correct, and the checking wing exists precisely because the gap between those two is invisible from the outside.

It will not rank the assistants yet. These tools change monthly, a ranking is stale before it is published, and a useful comparison needs weeks of real work in a real project.

It will not treat agents as solved. Where a floor describes something as working unattended, it will say on what kind of task and under what supervision, with the date.

And it will not carry affiliate links, including for the no-code platforms, which pay recurring commissions and dominate what currently ranks for these searches. What holds instead is simple: the checking wing applies to everything built with these tools, and no floor carries an affiliate link.