Problem
With several agents working in parallel, a terminal tells you what one agent said. It is harder to tell who owns the task, what changed and whether anyone reviewed it.
Technical decision
I put execution authority in the host process. The coordinator plans, the implementer edits and the reviewer returns a verdict. The desktop UI reads their state and exposes a bounded correction loop.
What I built
I built the desktop and backend workflow around those roles. An operator can move between agent windows and inspect a durable run history with file changes, findings and reported test results.
Evidence
In a saved development run, the team implemented a small formatting utility and its tests. The record included the plan, two changed files, a passing test command and reviewer approval. Other runs retained failure states, including a missing provider credential. This was a review of recorded behavior in the development version.
