Using Claude Code and alike with each agent working in a tab in the terminal feels very productive. You spend very little time waiting for AI since you basically start a task and go look at the next agents output.

But with this workflow for 6 months I can now see the downsides:

Finishing software is much harder

Working on 3-4-5 different features at the same time means you work in parrellel but you move slower on a single feature. It feels incredible moving on that many features, but it is harder to finish each. To finish a feature you need to deploy it and test it in the end, making sure it actually works.

I don't think we will see the exponitial boost from AI before AI can code itself.

AI is incredible useful, it is incredible at writing a lot of code with high accuracy, but when you use AI to code with, it gets the basic things incorrect.

Missing Plan mode

It is weird that gemini cli has no plan mode. In Claude Code you can do plan mode with two shift tabs, and it makes Claude present a plan after some investigation.

Less readable output in Gemini

Gemini

I am using Claude Code myself and there is people talking about that AI Coding agents just don't live up to standards, that it misses stuff and does not follow instructions, making it daunting to work with it. You can see it here AI Coding Sucks on youtube.

I can relate, using Claude Code now for 5 months, it does not sees to amaze how often it falls through and does simple things wrong.

Still helps

But I am using coding agents still because it helps me type less still, in programming when running commands from the commandline even 1 character wrong makes a mistake. This mistake the coding agent never makes, it does not mistype by one character.

Gemini Computer use just launched and I think it is going to be amazing. So much of programming is making integrations to different systems, and many systems does not have integrations yet.

Also integrations also requires something both searching, combining and inserting records, where a interface often combines those things into one.

So what is computer use? It is the next evolution of Large Language Models, it means that they are able to look at screenshots and take action. Previously LLM's could also look at screenshots (essentially just pictures) and try to predict what to do, but so much of automations is keeping a red thread through your actions. A model trained on a series of actions will be so much better at keeping the task in mind and trying different things while solving the task.

But we are still early!