Agent in the editor

Working in a codebase you own with an agent beside you in the editor or the terminal. You stay the author, and every line you accept is yours to explain.

Updated
Verified

Agent in the editor is the third rung, and the first where you hold the codebase. There is a repository, you work in it, and an agent works alongside you: suggesting code, editing several files at once, running the tests, explaining code you have not read before. You stay in the session while it works. You see each change as it is made, and you decide whether it stays.

Four categories sit here.

  • AI-assisted traditional coding, where you are the engineer and the model is a pair. You ask for a function, notice what it missed, and say so. Unglamorous, and still the most common way professionals work with these tools.
  • AI IDEs, where the editor indexes the whole codebase, makes edits across several files, and reads a rules file that carries your conventions from one session to the next.
  • Terminal and CLI agents, where you brief an agent from the shell and it works inside the repository: inspect the project, implement, run the tests, fix what broke.
  • Design to code, where a mockup or a design file becomes components. It sits here rather than at the base because its output lands in a real codebase that somebody has to maintain.

The work on this rung runs from you, to the agent, to the codebase, and back to you. That last step, back to you, is what separates it from the rungs below.

What this rung is for

Work in a codebase that has to last. Features, fixes and changes where you need to understand the result well enough to defend it in review and debug it during an incident.

It is also where most of the speed is safe to take, because you see every change before it lands.

Who it suits

  • Professional developers on codebases that will be maintained for years.
  • Teams who already review each other's code and want the same standard applied to an agent's.
  • Anyone stepping off the prompt-to-app rung who has decided to own the code. This rung asks you to read code. It is not where you learn to.

How to tell you are on this rung

  • There is a repository, it is yours, and you open it most days.
  • You read the agent's change before it lands, or you know you should.
  • You stay with the session while the agent works, rather than coming back later to a finished result.
  • You have caught yourself explaining the same convention to the agent twice.

If you hand over whole tasks and review the result afterwards, you are on the rung above.

What it costs you later

  • Every line you accept is yours. A model's code reads cleanly, which makes it easy to accept without reading. If you would not merge it from a junior engineer without review, do not merge it from a model. Understanding drops the longer review is skipped, and the bill tends to arrive months later, when nobody can explain the code that broke.
  • The context tax. What the agent knew in one session does not carry into the next unless it is written down. A rules file fixes that, and then needs maintaining like code. A rules file describing the conventions you wish you had is followed as faithfully as one describing the conventions you have. Context engineering for AI coding agents covers what belongs in one.
  • Drift and duplication. Left alone, generated code repeats itself and drifts from the patterns around it. Refactoring has to be scheduled on purpose, or it does not happen.
  • Design to code invents what is missing. With a design system in place it is strong. Without one, the tool invents one, and you inherit one-off components and values that nobody chose.
  • Conventions held by a product. The editors and agents on this rung change names, owners and pricing faster than codebases change. Keep conventions, specifications and rules in plain files in the repository, where any tool or person can read them, rather than only in one product's settings.

The step up

The rung above is Delegated execution, where you hand over a whole task and review what comes back. The step is worth taking when you can write a task down with a check that says it is done, and when your review of an agent's changes is reliable enough that you trust it without having watched them being made.

If neither is true yet, staying here is not falling behind. This rung is where most professional work with these tools happens.