bMAS: Blackboard LLM Multi-Agent System

• bMAS is a framework that integrates large language model agents with a shared blackboard to enable dynamic, context-aware problem solving.
• It employs a control unit to iteratively select specialized agent roles (planner, critic, cleaner) and form consensus, enhancing performance and token efficiency.
• Empirical results show bMAS outperforms static and dynamic MAS designs in tasks like commonsense reasoning and mathematics while reducing token costs.

A Blackboard-based LLM Multi-Agent System (bMAS) integrates the canonical blackboard problem-solving architecture with modern LLM agents to achieve dynamic, collaborative, and efficient problem solving. In this paradigm, a shared blackboard acts as the exclusive communication and memory substrate, enabling a diverse ensemble of agents—each possibly instantiated from LLMs of different capabilities and persona promptings—to iteratively cooperate, debate, and converge on solutions for complex queries. The approach is characterized by explicit shared context, dynamic agent selection, and iterative action rounds. Recent research demonstrates that bMAS can match or exceed the performance of both static and dynamically orchestrated LLM multi-agent systems in domains such as commonsense reasoning and mathematics, all while maintaining high token efficiency and adaptability (Han et al., 2 Jul 2025).

1. Blackboard Architecture and System Workflow

The core architectural feature of a bMAS is a globally accessible blackboard, which acts as the sole memory and message-passing hub for all agents. Every agent reads the entirety of the blackboard before acting, and all outputs—messages, plans, critiques, candidate answers—are posted as new blackboard entries. This design ensures every agent has access to the same context and history, eliminating the need for redundant private state or message duplication.

The problem-solving loop is structured as follows:

• Initialization: For each query $q$, the blackboard $B$ is initialized. An agent generation module AG uses $q$ and system-level instructions $I$ to produce a set of agents with roles and descriptions:

$\{(E_1, D_1), \ldots, (E_n, D_n)\} = AG(q, I)$

• Iterative Rounds: At each round, a control unit (represented as an LLM agent) selects a subset of agents for execution based on the current blackboard content and the query:

$$\{E_{i_1}, E_{i_2}, \ldots, E_{i_j}\} = ConU(q, B, \{D_1,\ldots,D_n\})$$

Each selected agent receives the entire blackboard as its prompt, producing a message $m_i$ appended to $B$.

• Consensus/Termination: This cycle repeats (with agent selection and execution dynamically adapting to the blackboard's evolving state) until a decider agent or majority consensus signals termination, after which the solution is extracted.

This iterative, context-driven orchestration allows the system to adapt collaboration patterns to each task instance, providing substantial flexibility when problem structures are ill-defined or evolve over time (Han et al., 2 Jul 2025).

2. Agent Roles, Communication, and Control

bMAS frameworks support highly diverse agent heterogeneity, including:

• Dynamically generated, query-specific experts: For each incoming query, the agent generation module synthesizes a set of domain-specific experts, each with a unique identity and description tailored to the problem.
• Predefined infrastructure agents: Common roles include:
  • Planner—decomposes complex queries into structured subplans.
  • Decider—signals when sufficient information is present for answer extraction.
  • Critic—examines the blackboard for errors or hallucinations, possibly prompting agent re-execution.
  • Conflict-Resolver—identifies and mediates contradictory agent outputs, potentially initiating secondary discussion threads.
  • Cleaner—prunes redundant or obsolete messages, maintaining token efficiency and board clarity.

All agent interaction is exclusively blackboard-mediated: agents neither message each other directly nor maintain private histories, enforcing strict global context. This aligns with classic blackboard architectures and contrasts with peer-to-peer or decentralized LLM-MAS paradigms (Han et al., 2 Jul 2025).

3. Action Selection, Execution, and Consensus

A distinctive operational aspect is action selection by a control unit agent, a meta-LLM that leverages the current state of $B$, the query $q$, and agent metadata to dynamically determine which agent(s) to execute per round. This enables:

• Selective engagement: Only agents relevant to the current state act, minimizing unnecessary computation and token expenditure.
• Context adaptation: If the problem evolves (e.g., new hypotheses, detected contradictions), the control unit may adapt agent selection on the fly.
• Consensus formation: Termination can occur via the decider agent's signal or by aggregating proposed answers using majority voting, where similarity functions (e.g., $V(a_i) = \sum_{j\neq i} \text{sim}(a_i, a_j)$) identifies the most converged solution.

This process ensures bMAS can fluidly adapt to both well-structured and highly dynamic problem domains without predefining strict workflows (Han et al., 2 Jul 2025).

4. Experimental Efficacy and Token Efficiency

bMAS has been empirically benchmarked against both static and dynamic MAS systems on commonsense and mathematics datasets (MMLU, ARC-Challenge, GPQA-Diamond, BBH, MATH, GSM8K) (Han et al., 2 Jul 2025). Key findings:

• Performance: bMAS achieved the best average results across all benchmarks, outperforming Chain-of-Thought (CoT) and static MAS by 4.33% and 5.02%, respectively, and matching or exceeding autonomous dynamic MAS baselines.
• Token Efficiency: The blackboard-centric approach reduces prompt length and token cost: as agents share a centralized message buffer, duplicate storage is avoided. The cleaner agent and judicious action selection further minimize token overhead. Table 3 in the paper quantifies this advantage, showing bMAS matches or improves on baselines at lower token budgets.
• Ablation Analysis: Disabling the control unit (i.e., reverting to simultaneous unregulated agent activation) increases average token cost and reduces signal-to-noise in blackboard content, underscoring the importance of adaptive agent orchestration and message cleaning.
• LLM Agnosticism: The framework generalizes across LLM backbones (demonstrated with Llama and Qwen), reinforcing architectural modularity.

5. Application Scenarios and Flexibility

The design is particularly well-suited for:

• Unstructured or evolving tasks: The blackboard’s shared, mutable context enables agents to integrate new evidence, revise plans, or resolve conflicts dynamically. This supports applications in open-ended research, scientific diagnostics, and real-time decision systems.
• Collaborative multi-agent planning in ill-defined environments: For clinical decision-making, integrated research, or multi-faceted evaluation, bMAS can handle dynamically shifting subgoals and agent roles, leveraging problem-adaptive agent activation.
• Token-sensitive settings: Scenarios constrained by LLM context window or inference cost particularly benefit from the blackboard’s compact, high-salience message maintenance.

6. Theoretical and Practical Considerations

bMAS advances LLM-MAS research by:

• Unifying collective memory: The blackboard enforces a single source of truth and prevents agent state divergence.
• Supporting dynamic coordination: Action selection and message content are adaptively contingent on the evolving problem state.
• Integrating specialized reasoning roles: The range of agent templates (planner, critic, cleaner, conflict-resolver) supports robust handling of hallucinations, contradictions, and subtask generation without manual scripting.

However, further research directions include:

• Enhancing agent diversity and specialization.
• Improving base LLM selection strategies and prompt tuning (e.g., incorporating MASRouter architectures).
• Deepening analysis of blackboard vs. alternative shared memory strategies.
• Scaling empirical validation to broader scenarios and real-world deployments.

7. Implications and Future Directions

The bMAS paradigm demonstrates that blackboard architectures—when combined with the capabilities of LLM agents and dynamic meta-control units—can drive state-of-the-art performance in collaborative AI reasoning, particularly in settings requiring flexible, adaptable, and token-efficient workflows. The blackboard’s collective memory model, dynamic agent selection, and centralized message discipline provide a robust foundation for complex and open-ended problem solving. The field is advancing toward more diverse agent ecosystems, increasingly sophisticated control logic, and generalized application across data modalities and domains (Han et al., 2 Jul 2025).

Future research will focus on increasing agent heterogeneity, enhancing control strategies, optimizing message curation, and evaluating deployments in progressively more complex, real-world settings.