An agent orchestrator, built and then replaced

The Challenge

Claude Code and Codex can each run several sessions at a time, but one terminal per agent stops working after two or three. In July 2026 I built UnifIDE to fix that: a local web app that runs many agent sessions in parallel across projects, each task in its own git worktree, with every diff reviewed before it merges. It shows across all projects which session is waiting on me, and a planning role locks a task's model, budget and scope before it starts. By September it stood at 349 commits and 123 test files.

The Solution

Then Claude added mods, which let the desktop app itself carry this kind of tooling, and I cancelled the Codex subscription that half of UnifIDE existed to support. Maintaining a second app around features the platform now had made no sense, so I set it aside. The part worth keeping was the planning: a project manager that keeps the plan, writes a brief per task and hands each brief to a worker. I rebuilt only that, as a mod, in three days. No worker starts until I have picked its model. Progress comes from workers reporting each finished step of their brief, so the bar never guesses. A finished worker shows as waiting for my review until its line in the plan is ticked. Live use on a game project turned up two bugs, a lost model choice and workers tagged with the wrong brief, and each was fixed with a test that fails without the fix. Around 1,400 lines and 32 tests. The picture is the mod's pane, drawn by its own code with sample workers.