Architecture - Capellaris

Architecture overview

Capellaris is built as a set of modules, so new agents and new system integrations can be added without touching the rest. Each component has one job and talks to the others through the same standard interface.

Architecture diagram

graph TB
    subgraph Interface["🖥️ User interface"]
        CHAT[Chat Platforms
Slack / Teams]
        API[API REST]
    end
    
    subgraph Core["⚙️ System core"]
        ROUTER[Message
router]
        CONFIRM[Human
confirmation]
        ORCH[Task
orchestrator]
    end
    
    subgraph Agents["🤖 Specialised agents"]
        MKT[Marketing]
        FIN[Finance]
        DEV[Engineering]
        HR[HR]
        SALES[Sales]
        OPS[Operations]
    end
    
    subgraph Learning["🧠 Learning system"]
        KB[Knowledge
base]
        ANALYZER[Pattern
analyser]
        OPTIMIZER[Prompt
optimiser]
    end
    
    CHAT --> ROUTER
    API --> ROUTER
    ROUTER --> CONFIRM
    CONFIRM --> ORCH
    ORCH --> MKT
    ORCH --> FIN
    ORCH --> DEV
    ORCH --> HR
    ORCH --> SALES
    ORCH --> OPS
    
    MKT --> KB
    FIN --> KB
    DEV --> KB
    KB --> ANALYZER
    ANALYZER --> OPTIMIZER
    OPTIMIZER --> ORCH
    
    style Interface fill:#e3f2fd
    style Core fill:#fff3e0
    style Agents fill:#e8f5e9
    style Learning fill:#f3e5f5

Modular

Independent components with well-defined interfaces, which is what makes the system maintainable and extensible.

Scalable

Designed to scale horizontally, supporting many agents and many concurrent tasks.

Extensible

New agents and new integrations can be added without modifying the core.

Main components

System core

The central brain that coordinates everything. It manages agent registration, orchestrates task execution and holds the system state.

Responsibilities:

Agent registration and management

Task coordination

Global state

Configuration management

Communication system

Handles message exchange between agents on Google's A2A (Agent-to-Agent) protocol, which keeps the communication standard and efficient.

sequenceDiagram
    participant U as User
    participant R as Router
    participant A1 as Agent 1
    participant A2 as Agent 2
    
    U->>R: Sends request
    R->>A1: Routes message
    A1->>A2: Asks for collaboration
    A2->>A1: Returns result
    A1->>R: Consolidated response
    R->>U: Delivers response

Human confirmation

The safety layer: every action an agent takes is approved by a human operator before it runs.

flowchart LR
    A[Task requested] --> B{Risk assessment}
    B -->|Low| C[Simple approval]
    B -->|Medium| D[Manager approval]
    B -->|High| E[Board approval]
    C --> F[Execution]
    D --> F
    E --> F
    F --> G[Report]

Technology stack

Python

Main language

FastAPI

Framework web

PostgreSQL

Database

OpenAI GPT

AI model

Backend

Redis for caching
RabbitMQ for messaging
Prometheus for monitoring

Integration

APIs REST e GraphQL
OAuth 2.0 e JWT
Webhooks for chat

System architecture