Overview
The Blackboard Pattern is a classic architecture from distributed AI, adapted for multi-agent LLM systems. Instead of direct agent-to-agent communication, agents interact through a shared data space.
Architecture
┌─────────────────────────────────────────┐
│ BLACKBOARD (Shared State) │
│ ┌─────────┐ ┌─────────┐ ┌─────────┐ │
│ │ Finding │ │Hypothesis│ │ Result │ │
│ │ #1 │ │ #1 │ │ #1 │ │
│ └─────────┘ └─────────┘ └─────────┘ │
└───────┬──────────┬──────────┬───────────┘
│ │ │
┌───┴───┐ ┌───┴───┐ ┌───┴───┐
│Agent A│ │Agent B│ │Agent C│
│(Read) │ │(Write)│ │(Both) │
└───────┘ └───────┘ └───────┘
Core Mechanics
Opportunistic Contribution
Agents monitor the blackboard for relevant information, process it when they have something to contribute, and post results back. No central coordinator assigns tasks.
Knowledge Types
- Hypotheses: Proposed solutions or interpretations
- Evidence: Facts, data points, observations
- Constraints: Rules that valid solutions must satisfy
- Results: Verified conclusions
Trigger Conditions
Each agent defines conditions for activation:
class ResearchAgent:
def should_activate(self, blackboard):
return (
blackboard.has_new("query") and
not blackboard.has("research_complete")
)
Benefits
Loose Coupling
Agents don't need to know about each other—only the blackboard schema.
Dynamic Participation
Agents can join or leave without disrupting the system.
Diverse Expertise
Different agent types contribute their specialized knowledge opportunistically.
Conflict Resolution
The blackboard can implement voting, confidence scoring, or expert arbitration for conflicting contributions.
Implementation Approaches
In-Memory Blackboard
Fast but not persistent. Suitable for single-session workflows.
Database-Backed
Persistent across sessions. Enables long-running collaborative investigations.
Event-Sourced
Every write is an append-only event. Full history and replay capability.
Use Cases
- Research synthesis: Multiple agents gather and analyze information
- Debugging complex systems: Agents contribute different diagnostic perspectives
- Creative brainstorming: Ideas build on each other organically
- Multi-perspective analysis: Different agent personas evaluate the same data
Modern Evolution: Arbiter Pattern
The Arbiter Pattern extends Blackboard with:
- Dynamic agent generation based on blackboard state
- Semantic task routing
- Self-organizing coordination
This enables mid-task adaptation and richer multi-agent dynamics.