Version Control Paradigm for AI-Assisted Coding
Source: Tweet thread Date: 2026-04-02
Summary
Half-formed but insightful tweet about how current version control (git) wasn’t designed for AI-assisted coding. Agent speedup is great for solo greenfield work but breaks down for collaborative development.
Key Claims
- Agent speedup is “great on greenfield/solo projects” but “when you start collaborating with other developers” the current VCS paradigm breaks.
- AI agents generate many small commits or large undifferentiated blobs — neither is ideal for code review.
- The missing primitive: something between “no history” (agent in agentic mode) and “full commit history” (human-mode).
- Proposed direction: “AI-aware” version control that understands agent vs. human changes, enables selective review.
Connection to Other Sources
Complements Cursor Scaling Agents (coordination failures in parallel agents) and cognitive debt concerns.
Concepts
- Coding Agents — VCS design for agent-assisted development
- Cognitive Debt — large undifferentiated AI commits are cognitive debt