AI Engineering

Every developer building with AI agents eventually hits the same wall. You write a masterpiece of a system prompt, loaded with rules, edge cases, and brand guidelines. It works perfectly... until you add one more rule. Suddenly, the agent gets "context fatigue," ignores half your instructions, and hallucinates a creative but entirely broken solution. We need to stop trying to make agents perfect on the first try. Instead, we suggest a new architectural approach: organizing ag

Ask a naive agent to find a bug in a codebase with thousands of repositories, and it will find the first textual match and stop. It won't ask why. It won't check whether what it found is actually relevant. It won't explore related services. It'll just return something and call it done. We, at Wix, discovered this the hard way. And it's what pushed us to build something more deliberate: a research engine that doesn't just search, but reasons about what it's looking for and why

When you have over 2,500 microservices and thousands of engineers working across a massive multi-product platform, "just use AI coding tools" is not a strategy. It's a starting point - and not a particularly useful one. The question we kept coming back to at Wix wasn't whether to adopt AI in our engineering workflow. That ship had sailed. The question was: what does agentic coding actually look like when you have strong domain boundaries between services, a monorepo with year

1. Introduction: The Challenge Understanding the engineering need for AirBot starts with recognizing the scale of the ecosystem it operates within. Wix operates a massive cloud-based development platform serving 250 million users. This generates over 4 billion HTTP transactions daily, feeding a data lake that currently holds 7 petabytes of data. To manage this volume, Wix Data Engineering maintains over 3,500 Apache Airflow pipelines (DAGs) . These pipelines handle everyth

If you use AI to code, you’ve felt the "Context Wall." You start a project, things go great, but as the codebase grows, the AI starts making suggestions that conflict with previous decisions. You spend more time correcting the model than writing code To solve this, I developed the Design-Log Methodology . It started as a way for me to remember my own decisions in large projects, but it has become the single most important tool in my AI-coding workflow. The AI is no longer jus