Agent Architecture

Torale uses a Gemini-powered monitoring agent (Pydantic AI) instead of the previous multi-provider pipeline.

Architecture Overview

APScheduler (cron trigger)
    ↓
Backend (job.py orchestrator)
    ↓
Agent (Gemini + Pydantic AI tools)
    ↓
Structured Response → Persist + Notify + Reschedule

Agent Workflow

The agent receives a prompt with the task's search query, condition, and previous evidence, then:

1. Gets current datetime via MCP tool
2. Retrieves memories from Mem0 (scoped to task)
3. Searches via Perplexity MCP tool
4. Evaluates condition against search results
5. Stores new meta-knowledge in Mem0
6. Returns structured response

Response Format

{
  "evidence": "Internal reasoning and findings (audit trail)",
  "sources": ["url1", "url2"],
  "confidence": 0-100,
  "next_run": "ISO timestamp or null",
  "notification": "User-facing message or null"
}

Backend Orchestration

job.py handles the execution lifecycle:

1. Creates execution record in database
2. Calls agent via A2A JSON-RPC
3. Maps response to task_executions row
4. Sends notifications if condition met
5. Auto-completes task if notify_behavior="once"
6. Dynamically reschedules via agent's next_run

Benefits

1. Single call — No multi-step pipeline coordination
2. Agent memory — Mem0 provides cross-run context without explicit state management
3. Dynamic scheduling — Agent adjusts check frequency based on context
4. Simpler error handling — Retry the whole call, not individual activities