Agent Ecosystem

The Agent Ecosystem and OAS

The OAS standard is not solely designed for evaluating exams. It is conceived as a structured Data Lake where a swarm of specialized artificial intelligences (each with a specific role) collaborates to autonomously create, update, and execute the educational lifecycle.

graph TD
    OAS[(OAS Specs Repository\nYAMLs)]
    
    Evaluator[Evaluator Agents\nEvaluate students against C0 Rubrics] -->|Lectura| OAS
    Curator[Curator Agents\nProcess PDFs/Docs to generate C0/C1/C2 Layers] -->|Lectura/Escritura| OAS
    Manager[Spec Managers\nOrchestrate and Compile Recipes] -->|Lectura| OAS
    Modernizer[Modernizer Agents\nMigrate old LMS packages to OAS] -->|Escritura| OAS
    ExerciseGen["Exercise Creation Agents\nGenerate dynamic exercises (C1)"] -->|Escritura| OAS
    ResourceGen["Learning Resources Agents\nSynthesize educational material (C2)"] -->|Escritura| OAS
    
    style OAS fill:#f9f,stroke:#333,stroke-width:2px

• Evaluator Agents (Evaluators): Consume the read-only standard. They cross-reference the ExerciseSpec with the student’s payload to issue precise mathematical scores and formative feedback.
• Curator Agents (Curators): Are responsible for ingesting raw knowledge. They analyze documents, laws in PDF, or raw texts and generate the different layers of the standard (extracting C0 taxonomies, designing C1 tasks, or preparing C2 context banks), saving the result as new YAML files.
• Modernizer Agents (Modernizers): Act as an ETL Pipeline for legacy systems. They ingest exported packages from traditional LMS platforms (Moodle, Canvas, Blackboard ZIP files, or formats like IMS QTI) and automatically transform them into the structural ecosystem of OAS.
• Exercise Creation Agents (Exercise Creators): Use didactic specifications and pre-existing contexts to iteratively generate new dynamic exercises aligned with the standards, saving them as structured instances in OAS.
• Learning Resources Agents (Learning Resource Generators): Responsible for distilling theoretical content and generating synthetic, adaptive, or pedagogical support educational material to feed the resource banks (C2).

Agent Network: Implementation in ColabEdu

Within the ColabEdu infrastructure, this ecosystem does not operate in isolation. We use a multi-agent approach where an orchestrator model routes requests to specialized Workers, ensuring optimal performance and context isolation.

graph LR
    User((Teacher / Student))
    
    subgraph ColabEdu Cloud
        Gateway[API Gateway]
        Supervisor[LangGraph4j Supervisor]
        
        Worker1[Evaluator Worker]
        Worker2[Curator Worker]
        Worker3[Rubric Generator]
        
        Supervisor -->|Delegates Tasks| Worker1
        Supervisor -->|Delegates Tasks| Worker2
        Supervisor -->|Delegates Tasks| Worker3
    end
    
    User <--> Gateway
    Gateway <--> Supervisor
    Worker1 -.-> SpecManager[Spec Manager API]
    Worker2 -.-> SpecManager
    Worker3 -.-> SpecManager
    
    style ColabEdu Cloud fill:#f0f8ff,stroke:#00509E,stroke-width:2px

The Supervisor (based on LangGraph4j) acts as the traffic brain. It analyzes the user’s intent (e.g., “Evaluate this History essay”) and exclusively awakens the Evaluator Worker with precise instructions. In turn, all these agents lack their own long-term memory; they use the SpecManager API as their single source of truth, ensuring that any change in law or rubric is immediately applied to all agents.