Latest updates

Zar started as a tool for web UIs. It now tests zero-native desktop apps as well, the kind built with Zig and WebKitGTK. Give it the repo and it compiles the app inside the sandbox, drives it, and records the run. No browser involved.

That changes where testing can happen. The app doesn't need to run on my machine or my screen. It builds and runs entirely inside the isolated sandbox, so a desktop app can be tested remotely, just like a website.

Everything else stays the same. A run still comes back as a chaptered video with pass or fail checks, a frame per step, and the errors it hit along the way. If it shows up on a screen, Zar can test it.

Zar can now run agent loops natively. I give it one job, an orchestrator agent plans the work and spins up worker agents to carry it out. I don't wire anything together myself anymore.

The orchestrator stays involved the whole way. It watches each worker's progress and sends corrections mid-run when one drifts. When the work comes back it reviews the result, and if something doesn't hold up, the worker goes back in. That loop repeats until the result is actually good.

The recording below is a real run. One orchestrator controlling a set of web agents, start to finish. I typed the prompt and let it go.