What is a Checkpoint?
Checkpoint: A core mechanism of Supervised Autonomy. It is a pre-defined point in an AI agent's workflow (like Cursor Agent or Windsurf) where the machine must stop, present its work or plan, and wait for a "Human Signature" before performing any irreversible actions.
Why Checkpoints are Mandatory
Without checkpoints, agents operate in Background Autonomy mode. This leads to Over-Delegation: an agent might make 50 incorrect commits overnight that you then have to manually unpick. Checkpoints ensure that mistakes are caught in Turn 1, rather than Turn 100.
Types of Checkpoints
1. Planning Checkpoint
The agent analyzes the task and proposes a "Plan of Action" (which files to touch, which patterns to use). The human reviews the plan for architectural soundness.
2. Modification Checkpoint
The agent proposes specific code changes. The human reads the diff and ensures no "Vibe-Code Smells" (like Magic Black Boxes) were introduced.
3. Destruction Checkpoint
Mandatory pause before any file deletion, dependency update, or schema modification.
Statistics
The "Human Signature"
A checkpoint is not just a "Continue" button. In the Clean Vibe methodology, it is a moment of Engineering Judgment. You are signing off on the AI's proposal as if you wrote it yourself.
Related Terms
- Human-in-the-Loop (HITL): The principle checkpoints implement.
- Supervised Execution: The mode that uses checkpoints.
- Over-Delegation: What checkpoints prevent.
Book Reference
- Chapter 14: Clean Agents — how to build supervised agent workflows.
- Chapter 18: The Manifesto — "AI Proposes. Humans Approve."