MCP for Stateful Agent Workspaces

AI21 Maestro is a test-time compute (TTC)-centric framework that runs multiple reasoning paths in parallel, but unlike read-only tasks, tasks that mutate state, like code, caused conflicts. When multiple subagents modified the same file or touched shared state, it led to change conflicts, file corruption, and rollback failures.

This problem wasn't limited to coding. Database agents need safe rollback of schema changes, document agents can overwrite each other's edits, and infrastructure agents can leave cloud resources half-configured. The key insight was that read-only agents can share an environment, but agents that write need isolation.

The solution was to add a workspace layer to MCP (Model Context Protocol). While MCP defines what a tool is and how to call it, it doesn't address where that tool executes or what state it shares. To address this, they defined 5 domain-agnostic primitives — initialize, clone, compare, merge, and delete — so that state-mutating agents can handle isolated environments under a common contract.

The coding agent implementation was built on git worktrees. Worktrees can be created quickly, each workspace corresponds to a branch, results can be naturally merged, and they share Git objects, keeping disk overhead low. This meant the protocol could handle "what needs to be done" while the implementation handled "how to isolate," and the two fit together well.

The benefits of this approach were clear.

• Parallel subagent execution: Each subagent works simultaneously in an independent workspace
• Speculative problem-solving: Multiple strategies run in parallel, and only the winner is merged
• Safe experimentation: Failed attempts can be instantly rolled back by deleting the workspace
• Domain-agnostic pattern: The same structure applies not just to code but also to databases, documents, and infrastructure

As a result, they were able to scale from 1 active approach to 16 parallel attempts without coordination overhead. The authors concluded that, unlike agents that read state, agents that mutate state require isolation as a protocol-level responsibility, and expressed hope that this pattern will be included in the official MCP protocol going forward.