Shared state, no drama: Scaling state-modifying agents with MCP workspaces

Scaling AI agents that modify shared state, such as a codebase, creates conflicts and data corruption when multiple agents act in parallel. The AI21 Maestro agentic framework encountered this problem, as its parallel reasoning approach broke down when agents needed to write changes. To address this, a "workspace layer" was proposed as an extension to the Model Context Protocol (MCP) to provide isolated environments for each agent. This layer introduces domain-agnostic primitives for creating, managing, and merging these isolated workspaces. For coding agents, this solution was implemented using git worktrees, which offer a fast and efficient way to create sandboxed environments for parallel development and testing.