Which AI Tools Do We Use?
Which AI Tools Do We Use?
The world of AI-assisted coding agents and tools is moving so fast that if you’re reading this in 2027, please move along unless you’re a computer science archaeologist. But if you’re reading this in early 2026, it might be interesting to see what tools Nitid finds valuable in January of 2026.
Although we don’t have any corporate standardization on what tools must be used, here’s an informal summary of what we’re currently using and how:
What Parts of Software Development Are We Letting AI Do?
Most of the time we’re using AI to generate the code and tests for individual stories/tickets. Usually these to-do items are chunked small enough that the approach hits the sweet spot. However, on some bigger stories, it makes sense to let AI do the planning for us. In those cases we’ll either have AI write a plan file in markdown, or we’ll have AI generate a full set of requirements, design, and tasks. Most of us prefer the former, but anything that lets us see the AI instructions, broken down into bite-size pieces, has been effective.
The actual writing of the code is an iterative process, with the AI doing the low-level code writing and the human doing the steering. We’ll often have to give the AI feedback on parts of the requirements that the AI incorrectly implemented, or instructions on how to better refactor the code. Although in our personal projects we may commit code unseen, in our professional work, we’re tightly reviewing every line the AI writes.
Once the tests pass, we manually check in the code to git. Although I’m willing to let AI commit directly to git on my personal projects, I like to take a little more control here in the office.
What’s Generating Our Code?
Most of us are Claude Code fans. The fact that it runs from a terminal still lets us feel “close to the metal” even when we aren’t writing any of the code.
Many of us are also VSCode fans, so Cursor, the AI-first fork of VSCode, is a popular choice. Although it’s not an efficient way to spend money/tokens on specific third-party models (like Claude), their in-house models are getting pretty good.
We also have folks running the official VSCode, using GitHub Copilot as their AI assistant.
And we have one intrepid explorer trying to manage multiple agents at once, each working on different feature branches. He’s checking out Sculptor which looks like a promising way to spin up parallel virtualized dev environments.
Any Other Tools?
We’re pretty interested in Steve Yegge’s Beads project that moves the AI “memory” from markdown files into git. Early playing with this product didn’t show much value, but months later, it’s worth another look!
And, on an upcoming project, we’ll be building an AI agent that will help automate tasks for overburdened medical caseworkers. Given our client’s existing HIPAA BAA with OpenAI, we’re looking at OpenAI’s agentic framework, AgentKit, but similar offerings from Anthropic and Google look capable too.