When a user hits a bug in Maestro, the last thing we want is for it to disappear into a support queue. We wanted a system where a user clicks “Report Issue,” and a few minutes later, a well-written GitHub issue appears in our repo – triaged, labeled, and ready for an engineer to pick up. No human in the loop unless the AI isn’t confident.
Here’s how we built it.
There’s a certain kind of confidence that comes from eating your own cooking. When we decided to rebuild a major subsystem of Maestro using Maestro itself, we weren’t trying to make a marketing point. We had a real deadline, a real feature spec, and a real question: could our own tool handle a non-trivial project from start to finish?
We wrote a specification for a new story management system — the part of Maestro that tracks which Coder is working on what, which stories are blocked, and how the Architect dispatches work. It was roughly 2,500 words of requirements written in plain English.
Most AI coding tools use a single model in a single conversation. You describe what you want, it writes the code, you copy it into your project. It’s convenient for small tasks, but it breaks down the moment your project has any real complexity.
The problem isn’t intelligence. The problem is that a single agent reviewing its own work has the same blind spots that produced the work in the first place.
Maestro’s offline mode wasn’t on our roadmap. It started with a phone call from a friend who does disaster relief logistics. He needed software tools in places where the internet is the first thing to go and the last thing to come back.
He was describing exactly the kind of work Maestro was built for — well-specified, practical software with clear requirements. The only problem: every AI development tool on the market required a persistent internet connection.
When people first see Maestro’s model configuration, they ask the same question: why would you use different models for different agents? Isn’t it simpler to just pick the best one and use it everywhere?
Simpler, yes. Better, no.
Every AI model has systematic biases in how it approaches problems. These aren’t bugs — they’re artifacts of different training data, architectures, and optimization targets. Claude tends toward careful, verbose implementations. ChatGPT tends toward concise, pragmatic solutions. Gemini often takes a different structural approach entirely.
Every AI coding tool claims to produce “production-ready” code. It’s become meaningless marketing language. So let’s be specific about what Maestro means when we say it, and what mechanisms actually enforce it.
Production-ready code, in our view, meets four criteria:
Most AI coding tools give you #1 on a good day. Maestro targets all four, structurally.
