AgentDiff — parallel plan review bench

Run several AI attempts at once, compare side by side, keep the winner and prune the rest

Problem

People coding with AI worry their own designs get mangled or commit history fills with slop. To test different directions they must create a branch or workspace per attempt, and managing and remembering those folders costs more effort than reviewing the code. Git assumes every edit deserves full traceability, which clashes with the flood of unstable experimental code in the AI era.

Solution

A lightweight workspace manager: launch several agent branches for one task, each producing a change snapshot, compare them in a side-by-side diff view (with AI summaries), then merge the pick into main and prune the rest without polluting git history.

Users

Indie developers coding with AI agents, tech leads at small teams, engineers who prototype frequently

MVP

Parallel attempt workspaces built on git worktrees; per-attempt diff snapshot with AI summary; side-by-side comparison view; one-click merge and prune; no extra commits

Revenue

Open-source core plus paid desktop/team edition (shared attempts, longer history), one-time or annual

Why now

Parallel multi-agent coding is becoming normal while version control still assumes stable code; this gap between git and agents is not owned by big vendors and is buildable solo.