Run a multi-agent build

The external MCP server acts on the most recently focused design. batch_design is atomic and undoable, but it does not lock nodes across agents. Assign one coordinator to review combined work.

1. Map pages and components with get_editor_state and batch_get.
2. Give each agent a distinct page, frame, component, or source file. Share exact IDs or paths. Each agent should use report_agent_activity with a stable ID and name, its current action, page/node or source path, and workspace (canvas, variables, or code). Use the same ID and name in batch_design.activity so Follow can track its streamed placements.
3. Have each agent finish canvas edits and let autosave complete before another edits the same source file. Do not mix canvas and source edits on the same nodes simultaneously. A dirty Code workspace draft locks canvas edits; save or discard it before handing the design to a canvas agent.
4. Build one component or section per batch so each operation appears on the canvas while the call runs. Verify each region with get_screenshot and inspect the Git diff for unexpected changes. A failed or interrupted batch rolls back its preview.
5. Clear activity with active: false when an agent finishes, then commit reviewed changes with ordinary Git.

Fanta's thirteen-tool external MCP surface reports agent activity and can follow an agent across canvas pages, component views, variables, and code files. It has no spawn_agent, vcs_begin_task, or cross-agent edit lock. Unrefreshed activity expires after about 30 seconds; parallel task assignment belongs to the client or team workflow.