After using AI tools and interacting with engineers who use them for almost a year now, a few patterns are starting to emerge from the field.

There are too many modes

• CoPilot has Ask, Edit, Inline Edit, and Agent
• Windsurf has Chat and Code
• Claude code has only one mode

We don't need so many. Likely these will be rationalized to one or two modes. I use only one mode and instruct via text if I want to ask a question and not make changes.

One textbox is enough for the most part.

Terminal based agents are refreshing

I used have Windsurf, CoPilot plugin from IntelliJ and VS Code. Recently I started using Claude code. I am using it far more than I use Windsurf now for prompting. Terminal feels far from accessible, fast, and consumes less resources. I suspect there are also benefits from integration between the AI model and the tool provider being the same.

Using AI from the terminal has a natural nudge for an engineer to use AI more than generating code. One tends to start using AI for managing file system, processes, analysing logs, and so on. It comes naturally.

Do not forget to write code by hand

Prompting agent to generate code, tends to build an inertia against writing code by hand. I notice that even for making simple code changes default behavior is to prompt. Prompting is suboptimal in such cases.

But I worry about something worse. Using AI is leading to proportional loss of confidence in writing code by hand.

I have written code most of my life. Now with AI writing so much of my code - the anxiety of having to write code by hand tends to creep in every so often. I don't like this. I don't want to lose the ability of being proficient in writing code by hand. Not just yet.

Refactoring as we knew it is dead. But Long live refactoring!

After performing deterministic IDE powered refactoring for almost 2 decades - I will admit I use it less and less - except maybe for rename.

But prompt based refactoring is not dead. In bad codebases which is overwhelming majority of useful codebases in the world - one can ignore refactoring only at their own peril.

Agent > Model > UI

There are several AI for SWE products available now. These products are available as IDEs, IDE plugins and node JS apps. These products embed the agents in them that interact with different AI models.

It is the agents that really differentiates these products from each other. At least at this stage of evolution.

If there are deficiencies in plugin/IDE one is willing to live with it. Similarly, on the other end the models are publicly available hence products allow switching between models. But the agents are the real heroes people are choosing products based on. Claude Code is so good that it has made IDEs a commodity.

Enterprises are facing a AI-SWE product choice dilemma

What I am noticing is that as businesses are choosing the product that they want to offer to their engineers - enterprise data protection is a big factor playing on their minds. But the fact that the "best tool" is changing so fast that they are falling behind. By the time these products are rolled out something much better emerges.

But at an individual level it is much easier. Choose monthly subscriptions instead of annual plans. Switch when something new emerges.

‍