Skip to content
Raindrop Developer Hub

Multi-Agentic System

The Multi-Agentic System pattern coordinates multiple specialized AI agents working together on complex tasks through shared memory and orchestrated workflows. This pattern uses collective intelligence where agents dynamically assume roles based on task requirements and learn from collaborative experiences.

Use this pattern when building:

  • Research and analysis workflows needing diverse expertise
  • Content creation pipelines with multiple review stages
  • Software development assistance with specialized roles
  • Complex decision-making processes spanning knowledge domains
  • Quality assurance systems with multiple validation perspectives
  • Multi-step problem-solving requiring specialized knowledge

Architecture Diagram

flowchart TB
    User[Complex Task Input]
    OrchService[Orchestration Service]


    User --> OrchService


    subgraph Orchestration ["Orchestration Layer"]
        OrchestratorAI["Orchestrator AI 70b+<br/>Task Planning & Coordination"]
        TaskDecomp[Task Decomposition]
        Validation[Result Validation]


        OrchService --> OrchestratorAI
        OrchestratorAI --> TaskDecomp
        OrchestratorAI --> Validation
    end


    subgraph Specialists ["Specialist Agents"]
        SpecialistService1["Specialist Service<br/>Dynamic Prompting"]
        SpecialistService2["Specialist Service<br/>Dynamic Prompting"]
        SpecialistServiceN["Specialist Service<br/>Dynamic Prompting"]
        SpecialistAI1["Specialist AI<br/>Configured Role"]
        SpecialistAI2["Specialist AI<br/>Configured Role"]
        SpecialistAIN["Specialist AI<br/>Configured Role"]


        TaskDecomp --> SpecialistService1
        TaskDecomp --> SpecialistService2
        TaskDecomp --> SpecialistServiceN
        SpecialistService1 --> SpecialistAI1
        SpecialistService2 --> SpecialistAI2
        SpecialistServiceN --> SpecialistAIN
    end


    subgraph Memory ["Shared Memory System"]
        SmartMemory[SmartMemory]
        Procedural["Procedural Memory<br/>Prompts & Protocols"]
        Episodic["Episodic Memory<br/>Workflow Learning"]
        Semantic["Semantic Memory<br/>Domain Knowledge"]
        Working["Working Memory<br/>Active Coordination"]


        SmartMemory --> Procedural
        SmartMemory --> Episodic
        SmartMemory --> Semantic
        SmartMemory --> Working
    end


    subgraph Communication ["Communication Layer"]
        Shared[Shared Utilities]
        TaskProtocols[Task Protocols]
        ResultAggregation[Result Aggregation]
        PromptGeneration[Dynamic Prompts]


        Shared --> TaskProtocols
        Shared --> ResultAggregation
        Shared --> PromptGeneration
    end


    OrchestratorAI <--> SmartMemory
    SpecialistAI1 <--> SmartMemory
    SpecialistAI2 <--> SmartMemory
    SpecialistAIN <--> SmartMemory


    Validation --> OrchService
    OrchService --> User

Components

  • AI (Orchestrator) - Large AI model (70b+ recommended) for task analysis, decomposition, planning, and validation
  • Service/AI (Specialist Pairing) - Configurable components that dynamically assume specialist roles through prompt injection
  • SmartMemory (Shared) - Centralized memory system enabling collaboration, learning, and context sharing
  • Shared - Common utilities for task serialization, result aggregation, and agent communication protocols

Logical Flow

  1. Task Reception & Analysis - Orchestration Service receives complex tasks and routes to orchestrator AI for requirements analysis

  2. Strategic Decomposition - Orchestrator AI analyzes complexity and decomposes work into specialized subtasks

  3. Agent Coordination - Orchestrator determines specialist configurations and coordinates task distribution through dynamic prompting

  4. Specialist Execution - Service/AI pairings receive dynamically generated role prompts and execute assigned subtasks

  5. Progress Monitoring - Orchestrator monitors specialist progress through SmartMemory, identifying dependencies and bottlenecks

  6. Validation & Integration - Orchestrator validates specialist outputs against criteria and coordinates revisions

  7. Result Synthesis - Orchestrator combines validated outputs into cohesive final results with consistency checks

  8. Collective Learning - All interactions and patterns stored in SmartMemory for system-wide learning and optimization

Implementation

  1. Deploy Orchestration Service - Configure main Service with orchestrator AI integration and task management

  2. Configure Orchestrator AI - Select and tune large model (70b+) with strong planning and validation capabilities

  3. Deploy Specialist Services - Create configurable Service/AI pairings with dynamic prompt injection capabilities

  4. Initialize SmartMemory - Set up shared memory with procedural templates, protocols, and domain knowledge

  5. Implement Shared Utilities - Develop common functions for task serialization, result aggregation, and prompt generation

  6. Production Setup - Add performance monitoring, workload balancing, failure recovery, and external knowledge integration

raindrop.manifest

raindrop.manifest
application "multi-agent-system" {


  service "orchestrator" {
  }


  service "specialist-pool" {
  }


  ai "orchestrator-brain" {
  }


  ai "specialist-brain" {
  }


  smartMemory "collective-memory" {
  }


  sql_database "coordination-db" {
  }
}

Best Practices

  • Use capable models - Deploy 70b+ parameter models for orchestration to ensure advanced planning capabilities
  • Design clear protocols - Establish standardized communication protocols and task handoff procedures
  • Implement validation layers - Build complete validation checking specialist outputs for consistency and quality
  • Plan for failure recovery - Design orchestrators to handle specialist failures, timeouts, and quality issues
  • Design flexible specialists - Create Service/AI pairings that can be dynamically configured rather than fixed-role
  • Optimize prompt libraries - Build complete libraries of role-specific prompts for dynamic combination
  • Implement context awareness - Ensure specialists understand their role in larger task and coordinate appropriately
  • Use procedural memory effectively - Store communication protocols and validation criteria for consistent behavior
  • Use episodic learning - Record successful collaboration patterns to improve future coordination
  • Organize semantic knowledge - Structure domain knowledge to support cross-agent knowledge sharing
  • Balance orchestrator load - Monitor orchestrator decision-making load and scale or optimize as needed
  • Optimize specialist utilization - Track usage patterns and adjust scaling policies to match demand