NOIR

Switchboard is shared decision infrastructure for multi-agent systems. It helps AI assistants reuse a user’s decisions within explicit boundaries, track which work depends on those decisions, and identify what needs review when the user changes their mind. As people delegate work across multiple assistants, they become responsible for carrying preferences, permissions and corrections between them. Switchboard gives compatible agents a common decision service through MCP. Each request is checked against scoped, versioned rules, hard locks and review requirements. Eligible answers can be reused; unresolved questions reach the user through one operator window or a phone-friendly Focus mode. Its central technical contribution is connecting decision reuse with correction impact. Receipts link answers to rule versions, while recorded dependencies identify affected drafts and tasks. Before applying a correction, users can preview that impact. A recalibration hold pauses reuse of the changed rule while unrelated rules remain available. Users can also steer proposals, inspect revisions and approve them separately. Meta Muse Spark Open, accessed through Baseten, provides optional semantic matching, steering suggestions and revision summaries. The model selects among eligible rules while application logic enforces scope and authority. Modal hosts the shared service and a separate public playground with disposable visitor workspaces. Switchboard supports integrations for Muse Code, Codex, Claude Code, OpenCode, Hermes, Claude Desktop, ChatGPT and Claude, with enforcement varying by harness. The submission includes a working deployment, public sandbox, three-minute walkthrough, 244 passing backend tests and a reproducible synthetic benchmark showing the tradeoff between incorrect automatic decisions and human review.