Domain Agents

KULVEX uses a modular agent architecture. Each agent owns a specific domain (weather, home, security, etc.) with its own pattern matching, tools, and response templates.

How It Works

User Message
    │
    ▼
  AgentRouter.detect_intent(message)
    │
    ├── Match found → Agent.handle() → ExecutionResult
    │     ├── Template response (zero LLM, instant)
    │     └── Tool execution + synthesis (minimal LLM)
    │
    └── No match → Agentic Chat (multi-round tool_use via LLM)

Agent Routing

The AgentRouter checks each agent in priority order. Each agent has:

Priority — Lower number = checked first (10-90)
Patterns — Regex patterns that match user messages
Tools — Available tools for that domain
Templates — Zero-LLM responses for simple queries

When a message matches an agent’s patterns, that agent handles the request. If no agent matches, the message goes to agentic chat for general-purpose reasoning.

API

# List all agents
curl http://localhost:9100/api/ai/agents
 
# Test intent detection
curl "http://localhost:9100/api/ai/agents/detect?message=what's the weather"

Voice Pipeline Agent List