AI is redefining what it means to be a web engineer.
I’m building with it and writing about what’s actually working.
I’m Steven Sacks, a fullstack engineer based in Tokyo, and I’m all-in on Agentic Engineering. I’m more engaged than I’ve been in years.
What I’m figuring out lands here in the articles I’m writing, and in GAIA, my open-source Claude Code workflow for web projects.
I write in English, and every article is translated into Japanese for my community of engineers in Japan, where I call home.
What I believe
Working with AI well is a teaching relationship that runs both ways.
The engineer sets the spec, writes the plan, defines the tasks, and corrects the drift. AI executes. But every correction shapes the next interaction, and every gap it surfaces teaches you what you left out. Neither one has the full picture alone.
Most bad AI output is an input problem.
A model can only work with what you hand it. When the result is wrong, the fix is usually sharper explanation or a smaller task. Good input is your responsibility. So is knowing when the model has hit its ceiling.
Autonomous AI workflows are only as good as their gates.
Models can verify the work of other models, and that helps. But AI drifts in ways code does not and humans do not. It has its own category of failure. Guardrails reduce it. Running them more than once catches what a single pass misses. How much autonomy you hand over is really a question of what oversight matches the risk you are willing to own.
