Why We're Building an AI Software Engineering Environment That Shows Its Work

AI coding tools are good at producing code. They're much worse at showing you why you should trust it.

You ask for a change, you get a diff, and then you're left with the real job: reading it, running the tests, and deciding whether it actually works. The proof is your problem.

That's the gap we're working on at Zylvex Technologies, and it's why we're building ZylCode.

The idea: evidence, not just output

ZylCode is an engineering workspace built around one rule: a result isn't done until there's evidence for it.

Instead of a chat box that returns code, you describe an outcome and ZylCode turns it into a mission:

1. Intent: you describe what should change.
2. Plan: the work is broken into explicit steps.
3. Evidence: candidate changes are checked by running your real test suite.
4. Verified result: you see what passed, what didn't, and why.

A few design decisions

Nothing runs by itself. Queued work executes only when you press Run queued. The agent doesn't get to quietly act in the background.

Commands run for real. The built-in terminal runs commands in your actual shell, so what you see is what happened. To be upfront: it's a command runner, not a full interactive PTY, and the working directory persists per session.

Verification is part of the workflow. Missions come with an explicit verification path from the start, not as an afterthought.

Where it stands today

ZylCode is a Developer Preview, and we'd rather be honest about that than oversell it.

• Available: missions (the intent pipeline) and repository intelligence
• Limited: the runtime for web/desktop
• Coming soon: a design studio and a multi-agent board

See it running

This is the real application, not a mockup:

Try it and tell us what breaks

The code is here: github.com/zylvex-tech/zylcode

We'd really like feedback on one question: what evidence would you need before you'd trust an AI-generated change enough to merge it? Tell us in the comments.