Core Concepts

Agents

An agent is an autonomous reasoning unit powered by a large language model. Each agent has a defined role, memory scope, tool access, and communication channels. Agents operate asynchronously and coordinate through the neural router.

Agent definition schema:

# agents/my-agent.yml
name: cogniscan-Ω
role: market-intelligence
model:
  provider: openai
  name: gpt-4o
  temperature: 0.2
memory:
  type: hybrid
  short_term: 1000  # context window
  long_term: vector  # pgvector-backed
tools:
  - web-search
  - data-analysis
  - slack-notify
channels:
  - type: graphql
    endpoint: /v1/agents/cogniscan
  - type: event-bus
    topics: [market-deltas, competitor-alerts]

Neural Router

The neural router is the central coordination layer. It receives tasks, determines which agent(s) should handle them, and routes the workload with optimal latency. Routing decisions use a combination of semantic matching, agent load, and historical performance.

# config/routing.yml
routing:
  strategy: semantic-hybrid
  matchers:
    - type: embedding
      model: text-embedding-3-large
      threshold: 0.92
    - type: rule
      patterns:
        - "refund" -> agent: support-triage
        - "deploy" -> agent: devops-core
        - "forecast" -> agent: cogniscan-Ω
  fallback:
    strategy: round-robin
    pool: general-purpose-agents

Consensus Protocol

For critical decisions, Departments uses a multi-agent consensus protocol. When a decision threshold is crossed, the router spawns a consensus round where multiple agents independently evaluate the same input and vote.

# Trigger a consensus round via the API
curl -X POST https://api.departments.cyrus365.com/v1/consensus \
  -H "Authorization: Bearer $DEP_TOKEN" \
  -d '{
    "task": "evaluate-acquisition-target",
    "mode": "super-majority",
    "context": { "company": "Nexigen AI", "valuation": 420000000 },
    "agents": ["cogniscan-Ω", "aethra", "synthmind"],
    "timeout_ms": 5000
  }'

Nexus Architecture

A Nexus is a deployed instance of the Departments system. It comprises one or more clusters, each containing agents, routers, and storage nodes. The architecture is horizontally scalable and supports multi-region deployment.

# nexus.json
{
  "version": "4.2.7",
  "name": "production-nexus",
  "clusters": [
    {
      "id": "us-neural-1",
      "region": "us-east-1",
      "agents": ["cogniscan-Ω", "nexusflow", "aethra"],
      "replicas": 3,
      "storage": {
        "type": "pgvector",
        "connection_pool": 20
      }
    },
    {
      "id": "eu-cortex-1",
      "region": "eu-west-1",
      "agents": ["nexusflow", "synthmind"],
      "replicas": 2,
      "storage": { "type": "pgvector" }
    }
  ],
  "global_routing": {
    "type": "geo-proximity",
    "fallback_region": "us-neural-1"
  }
}

Memory System

Departments uses a tiered memory architecture to give agents both immediate context and persistent recall. Short-term memory resides in the agent context window, while long-term memory is stored in vector databases and retrieved via semantic search.

Architecture Diagram

Below is a simplified representation of data flow through a Departments nexus:

  ┌─────────────────────────────────────────────────────────┐
  │                    API Gateway                            │
  │  GraphQL ←→ REST ←→ WebSocket ←→ Event Bus               │
  └──────────────┬──────────────────────────────────────────┘
                 │
  ┌──────────────▼──────────────────────────────────────────┐
  │                Neural Router                             │
  │  [ semantic matcher ] [ load balancer ] [ fallback ]    │
  └──┬────────────┬────────────┬──────────────────┬─────────┘
     │            │            │                  │
  ┌──▼───┐   ┌────▼────┐  ┌───▼────┐      ┌─────▼──────┐
  │Agent │   │ Agent   │  │ Agent  │      │ Consensus  │
  │Pool 1│   │ Pool 2  │  │ Pool 3 │      │ Engine     │
  └──┬───┘   └────┬────┘  └───┬────┘      └─────┬──────┘
     │            │            │                  │
  ┌──▼────────────▼────────────▼──────────────────▼───────┐
  │               Memory Layer (pgvector + Redis)          │
  └───────────────────────────────────────────────────────┘