Coding agents make an industry out of the smallest task

Coding agents are making an absolute industry out of the smallest task.

I asked for hold-Space dictation in my coding-agent setup. Hold Space, speak, release, and the transcript appears in the editor. That’s the feature.

The first implementation arrived with a package, a setup flow, secret handling, documentation, tests and a lot of code.

It also deleted the text already in the editor.

The order was wrong

None of the supporting work was inherently wrong. The order was.

When I pressed the agent, it said it wanted to make sure the feature was production ready. It had started designing for every imagined production concern before proving the basic interaction worked.

This is the bit I’m getting tired of. Generated code is cheap, so models keep producing more of it. Every extra file and abstraction adds more for me to reason about, and more places for the actual problem to hide.

Smallest useful path first

I’ve started pushing back with a simpler rule:

1. Get the smallest useful path working first.
2. Prove it in the real interface.
3. Then add the structure the product actually needs.

That is not an argument against tests or production standards. On a change that touches data, authentication, money or infrastructure, the safety work can be most of the job, and it should be. It is an argument for sequencing: the structure should match what you’ve actually proved.

Production-ready code that doesn’t work is not production ready.

What a team gets from this

Agents will happily fill a pull request with plausible scaffolding. Part of working with them well is knowing when to say “smaller”, and having the checks in place so the small version is still a correct one. That’s a large part of how I use them on client work.

If your team’s agents are producing more code than product, let’s talk.