Field note
Local coding agents are becoming shared workspaces whose main product is coordination state, not another model wrapper. The category is visible now because four independent projects are solving coordination failures that appear after teams move beyond one isolated agent session.
Munder Difflin exposes a local multi-agent office with coordination, mailbox, routing, and shared state. Concord MCP adds presence, communication, ownership, and handoff state across multiple coding-agent tools. Browser Use's macOS Harness adds a managed desktop execution surface, while OpenClaw 2 packages browser, messaging, memory, skills, models, and automations into one runtime.
The workspace layer assigns ownership, records presence, carries messages, preserves handoff context, and gives agents managed execution surfaces around the same repository. That mechanism matters because parallel agents fail less often from an inability to generate code than from duplicated work, conflicting edits, unclear responsibility, and context lost between sessions.
Teams should evaluate agent workspaces by conflict prevention, recoverable handoffs, observable ownership, and state continuity rather than by the number of agents they can launch. This adds concrete implementation evidence to the New Runtime harness pattern.
These are project capabilities and design directions, not comparative evidence that one workspace produces better software or safer coordination than another. Watch whether the tools converge on interoperable ownership and handoff contracts, or remain separate shells tied to their own agents and local state.