LbMAS: Blackboard-Driven Multi-Agent Intelligence
This presentation explores LbMAS, a novel multi-agent architecture for large language models that uses a blackboard system to coordinate specialized agents. Through dynamic agent selection, shared memory, and iterative consensus mechanisms, LbMAS achieves superior reasoning performance while dramatically reducing token costs compared to traditional multi-agent approaches.Script
In most multi-agent AI systems, each agent maintains its own memory, leading to redundant context and skyrocketing token costs. LbMAS flips this model entirely: all agents share a single blackboard, a public memory where every insight, critique, and decision is visible to the collective.
This blackboard is not just storage. It's the conversation itself. When a planner posts a strategy, when a critic identifies a flaw, when an expert contributes domain knowledge, it all accumulates in one place. This centralized design cuts token consumption dramatically because no agent needs to reconstruct the full dialogue independently.
How does the system decide which agents to activate at each step?
The control unit reads both the user query and the current blackboard, then selects which agents to activate for that round. Complex problems trigger critics and conflict resolvers. Simple queries might converge in a single planning step. This adaptive orchestration means the system self-tunes its reasoning depth to match problem complexity.
Rather than trusting a single agent's output, LbMAS computes consensus. Each agent proposes an answer, and similarity scores reveal which solution the ensemble collectively supports. This voting mechanism filters out outlier errors and amplifies the wisdom of the agent crowd.
On challenging reasoning benchmarks, LbMAS outperformed both chain of thought and static multi-agent baselines by 4 to 5 percentage points. More striking is the token efficiency: it achieved 72.6% accuracy on the MATH dataset while consuming just 4.72 million tokens, far fewer than competing systems that require expensive workflow optimization phases. The system adapts online, in real time, without pretraining overhead.
LbMAS demonstrates that intelligent coordination through shared memory and dynamic agent selection can unlock both stronger reasoning and radical cost efficiency. The blackboard is not just a metaphor; it is the operational heart of scalable multi-agent intelligence. Visit EmergentMind.com to explore more research and create your own videos.
