How I Build AI Agent Systems at Rocket.new (From the Inside)

Source: Dev.to

I’ve been building developer tools for five years.

For the first three years at DhiWise, we automated one thing: turning Figma designs into production code. Pick a framework (Flutter, Kotlin, React), upload your design, get clean code out. That was the whole product.

Then everything changed.

The Pivot That Changed How I Build

In 2023 we launched WiseGPT – a context‑aware AI coding assistant that plugged into VS Code and understood your local codebase. No prompt engineering. Just describe what you want, and it generates code that fits your actual project structure.

That was my first real exposure to building AI agent systems in production. Not demos. Not prototypes. Real agents, serving real developers, at scale.

By 2025, the company rebranded to Rocket.new and the product became something much bigger: an AI‑first application builder that generates complete, full‑stack, production‑ready applications from plain English prompts. I’m one of the engineers building the agent systems behind that.

What an AI Agent System Actually Looks Like in Production

A lot of tutorials show a single LLM call: you send a prompt, you get a response. That’s not an agent system; that’s autocomplete.

At Rocket.new, a single user prompt like “build me a SaaS landing page with a waitlist form” triggers a coordinated pipeline of specialized agents:

• Research agent – analyzes the request and identifies what components are needed.
• Design agent – makes UI/UX decisions.
• Code generation agent – writes frontend, backend, and database logic.
• Validation agent – checks the output for errors before it reaches the user.

These agents don’t run strictly sequentially. They run in parallel where possible, coordinate decisions through a shared context layer, and hand off state to each other asynchronously.

The Hardest Problem: Context Engineering

The biggest technical challenge isn’t the LLM calls; it’s context engineering – deciding what information each agent needs to do its job well, and making sure that information is available at the right time in the right format.

• Too much context and the model gets confused.
• Too little and it makes wrong assumptions.

The art is in building pipelines that route exactly the right context to each agent at each step. This is what I spend most of my time on at Rocket.new.

What I’ve Learned Building This

1. Hybrid architectures beat pure LLM systems
   Combining utility‑based decision making (deterministic rules) with LLM‑based reasoning gives you systems that are both reliable and flexible. Pure LLM systems drift; hybrid systems are controllable.

2. Error cascading is the real enemy
   When one agent makes a small mistake, the next agent inherits that mistake and builds on it. By the time you’re three agents deep, you have a completely wrong output. The solution is validation gates between agents, not just at the end.

3. Latency matters more than accuracy at first
   A faster, slightly less perfect answer beats a slow, perfect one. Users lose trust when they wait. Build for speed first, then improve accuracy incrementally.

The platform now serves 400,000+ developers across 180 countries and has generated 500,000+ production‑ready applications.

If you’re building AI agent systems and want to trade notes, I’m always open to it.

Find my other technical writing at .