IoT Scenarios for Hercules Micro-Agents

Hercules is built as a tiny, self-improving micro-agent. One of its most compelling deployment targets is Raspberry Pi at the edge: a low-cost device that runs a single-purpose agent, connects to sensors and actuators, and participates in a larger agent mesh when needed. The LLM itself runs externally (YandexGPT, Ollama Cloud, or any OpenAI-compatible provider) — the Pi is the agent runtime, not the model host.

This document collects realistic micro-scenarios for B2C and B2B use. Each scenario is deliberately small, vertically focused, and deployable as one Hercules agent on one Pi.

1. Reference architecture: one agent on one Pi

┌───────────────────────────────────────────────┐
│              Raspberry Pi (edge)              │
│  ┌────────────┐  ┌──────────┐  ┌───────────┐ │
│  │  Sensors   │  │ Actuators│  │  Hercules │ │
│  │  GPIO/I2C  │  │  GPIO    │  │  micro-   │ │
│  │  MQTT      │  │  relay   │  │  agent    │ │
│  └─────┬──────┘  └────┬─────┘  └─────┬─────┘ │
│        │              │               │       │
│        └──────────────┴───────────────┘       │
│                       │                       │
│              ┌────────▼────────┐              │
│              │  skills/ memory │              │
│              │  data/ (local)  │              │
│              └─────────────────┘              │
│                       │                       │
│              ┌────────▼────────┐              │
│              │   LLM provider  │              │
│              │  (cloud/fallback)│             │
│              └─────────────────┘              │
└───────────────────────────────────────────────┘

Key assumptions for this phase:

2. B2C scenarios

2.1 Smart greenhouse for home gardeners

Goal: keep one balcony greenhouse or small grow tent in ideal condition.

Agent name: hercules-greenhouse-{id}

Devices on Pi:

Skills:

Memory:

Why micro: one Pi, one greenhouse, one agent. The user asks via Telegram: “How is my basil?” — agent reads sensors, consults memory, and answers with context.

2.2 Home energy monitor

Goal: track electricity usage of one appliance or one room and give advice.

Agent name: hercules-energy-{id}

Devices on Pi:

Skills:

Memory:

Why micro: cheap to deploy per room. Over time skills improve to distinguish appliances by power signature.

2.3 Baby room / elderly care monitor

Goal: watch comfort and safety of one room, not a whole house.

Agent name: hercules-room-{id}

Devices on Pi:

Skills:

Memory:

Why micro: privacy-first, one room, local audio triggers only. LLM is used for natural-language summaries, not for streaming audio.

2.4 Pet feeder and tracker

Goal: feed a pet on schedule and answer owner questions about feeding history.

Agent name: hercules-pet-{id}

Devices on Pi:

Skills:

Memory:

Why micro: one pet, one device, one agent. Skills evolve around this single pet’s routine.

2.5 Workshop air quality guard

Goal: protect a single room (garage, workshop, 3D-print room) from harmful air.

Agent name: hercules-air-{id}

Devices on Pi:

Skills:

Memory:

Why micro: targeted at one hazardous room, not whole-building HVAC.

3. B2B scenarios

3.1 Cold-chain spot checker (small shops, pharmacies, food trucks)

Goal: ensure one fridge or freezer stays in temperature range and logs compliance.

Agent name: hercules-coldchain-{id}

Devices on Pi:

Skills:

Memory:

Why micro: one fridge = one agent = one invoice. Skills improve per product type across deployments.

3.2 Office desk / meeting room comfort agent

Goal: optimize one meeting room or open-space zone.

Agent name: hercules-room-comfort-{id}

Devices on Pi:

Skills:

Memory:

Why micro: one room, one agent, easy to scale floor-by-floor.

3.3 Small-server / network closet watchdog

Goal: monitor one rack, closet, or edge cabinet and react to environmental issues.

Agent name: hercules-closet-{id}

Devices on Pi:

Skills:

Memory:

Why micro: one cabinet, one agent. IT can deploy dozens with identical base config and per-site overrides.

3.4 Vending machine / locker assistant

Goal: make one unattended machine conversational and proactive.

Agent name: hercules-vending-{id}

Devices on Pi:

Skills:

Memory:

Why micro: one machine, one agent, independent of central platform.

3.5 Construction / rental equipment telemetry

Goal: track usage and health of one generator, compressor, or scaffold set.

Agent name: hercules-equipment-{id}

Devices on Pi:

Skills:

Memory:

Why micro: one asset, one agent, survives intermittent connectivity by buffering data locally.

4. Common design patterns across scenarios

Pattern How Hercules helps
One agent = one physical thing Simple deployment, simple ownership, simple billing.
Local skills, cloud LLM Pi handles logic; LLM gives natural language and reasoning.
Skill improves per deployment Each device learns its own thresholds and owner preferences.
Mesh when needed A “home hub” or “fleet manager” agent can ask a room agent for status.
Offline resilience Sensors keep logging; alerts queue; skills do not depend on constant cloud.

5. Hardware baseline

A minimal reference Pi setup for any scenario:

Hercules itself remains small: one process, one data/ folder, one config file.

6. From scenario to product

For each scenario above, the path to a sellable product looks the same:

  1. Agent template — pre-built skills, memory schema, and device wiring for the scenario.
  2. SD card image — ready-to-flash Pi image with Hercules + template.
  3. Activation flow — user sets Wi-Fi and API key, agent registers in mesh/cloud.
  4. Managed mesh — optional central dashboard lists all deployed agents.

The same runtime serves B2C (one device at home) and B2B (hundreds of devices in field). The only difference is scale and how agents are provisioned.

7. Next steps

For the mesh vision, see docs/AGENT-MESH-EN.md.
For the phased roadmap, see docs/ROADMAP-EN.md.