Home Assistant Autonomous AI Agent

A proactive, LLM-powered smart home agent that runs entirely within Home Assistant's python_script sandbox. It analyzes context (weather, calendar, sensors), maintains memory, and autonomously controls devices or sends alerts.

Features

• Autonomous Loop: Analyzes data and executes multi-step actions without user intervention.
• Tool System: Can check S-Bahn times, weather, read sensor states, and control devices.
• Memory: Persists user facts and corrections via template sensors (long-term learning).
• Proactive: Triggers urgent alerts (e.g., high CO2) or helpful reminders based on context.
• Sandboxed: Runs without external Python imports, fully compatible with HA's restricted environment.

Requirements

• Home Assistant with python_script integration enabled.
• LLM Integration: An entity (e.g., ai_task.qwen3_instruct) provided by an integration like ollama or ai_task.
• Template Sensors: Required for storing agent memory and state.

Installation

1. Enable Python Scripts: Add python_script: to your configuration.yaml and restart HA.

2. Deploy Agent: Copy agent.py to your /config/python_scripts/ folder.

   • Configuration: Edit the PERSONAL CONFIGURATION section in agent.py (User Name, Language, Entity IDs).

3. Setup Memory & State: Copy the template sensors from config.yaml into your HA configuration.yaml (under template:). These store the agent's "Learned Facts" and history.

4. Automation: Use automation.yaml as a template. It handles:

   • Waking up the local PC (if hosting the LLM).
   • Triggering python_script.agent on a schedule.
   • Shutting down the PC after completion.

Usage

Manual Trigger: Call the service in Developer Tools:

service: python_script.agent
data:
  max_iterations: 10
  entity_memory: sensor.ai_memory_store
  llm_entity: ai_task.qwen3_instruct

Logic Flow:

1. Gather Context: Reads weather, calendar, and memory.
2. Think: Sends prompt to LLM.
3. Act: LLM chooses a tool (e.g., CONTROL_DEVICE, GET_SBAHN).
4. Repeat: Loops until task is finished or max iterations reached.