The shared blackboard pattern is back and it's solving multi-agent context loss

Описание и применение

Шаблон совместной доски (blackboard) успешно решает потерю контекста в многопользовательских системах LLM, повышая точность и эффективность, но требует строгой проекции и управления.

Задача

Agents in a multi‑agent system were unable to see shared context, causing repetitive loops and loss of information across handoffs.

Результат

The blackboard pattern improved information fidelity at the coder stage by 62%, correct file targeting by 27.5 percentage points, and achieved 100% ledger accuracy versus 0% in a production system, with 26× lower latency and 29× lower token consumption.

Рабочий процесс

• 1. Agents publish messages with optional references to prior messages.
• 2. Agents search the board by content, receiving at most k relevant messages.
• 3. Agents visit specific messages by address.
• 4. A Supervisor LLM orchestrates agent turns based on board state.
• 5. Governance policies enforce scoped retrieval, temporal supersession, provenance tracking, and policy‑governed propagation.

Доказательства

The shared blackboard wasn't shared.

Источник

Information Fidelity Score at the Coder stage: 0.586 vs. 0.362 — a 62% improvement

Источник

100% ledger accuracy versus 0% for the debate panel.

Источник