Agent Scratchpad Workspace Pattern

Overview

An isolated, agent-owned workspace pattern with a three-part directory structure: transient sessions, read-only context, and pending artifacts. Implements human-in-the-loop validation before artifacts merge into the main codebase.

Structure

agent-scratchpad/
  sessions/    # Transient agent work (ephemeral, auto-cleaned)
  context/     # Read-only reference material (injected before execution)
  artifacts/   # Pending outputs awaiting human approval

Key Principles

• Isolation: Agent work happens in a dedicated space, never directly in production code
• Read-only context: Agents receive curated context but cannot modify source material
• Artifact staging: All agent outputs are pending until human approval
• Session ephemerality: Working state is transient; only approved artifacts persist

See Also

• ai-agent-execution-constraints - Complementary constraints for agent execution boundaries
• context-refresh-session-state - Session state management patterns for AI agents
• boundary-guard-hooks - File access enforcement for agent workspaces