{
  "id": 4420814,
  "title": "Building Full Shelf: An Agent Fleet for Food-Bank Operations",
  "url": "https://urgent.news/2026/08/30/building-full-shelf-an-agent-fleet-for-food-bank-operations",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-30T12:14:58.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/quietmoose/building-full-shelf-an-agent-fleet-for-food-bank-operations-2cep"
  },
  "original_language": "en",
  "account": "The AllThingsAgenticHackathon at Google Cloud showcased an innovative AI solution called Full Shelf, designed specifically for managing the complex operations of food banks. This agentic control plane addresses numerous challenges that arise when dealing with perishable food items, such as inventory management, refrigerated transportation, safety regulations, and community commitments.\n\nFull Shelf creates a centralized dashboard for the entire operating day, aggregating vital information such as inventory, vehicles, deliveries, partner partnerships, custody, incidents, approvals, and pending tasks. It then utilizes a fleet of five specialized agents, powered by Google's Gemini 3.5 Flash and Vertex AI, to evaluate the available food, vehicle capacity, delivery obligations, safety constraints, and partner options. These agents generate a feasible operating plan, adapting to any changes in real-time.\n\nWhen a food recall occurs, Full Shelf quickly screens the notice, extracts the affected product and lot, traces current custody, halts unsafe inventory, and determines what needs to be replaced. Unconfirmed inventory, missing evidence, and unresolved shortages are preserved, ensuring transparency throughout the process. A crucial rule implemented in the system is: Gemini recommends, a human approves, and only the ledger writes. This ensures that the AI agents make recommendations, but a human director reviews and validates the changes before any action is taken.\n\nRather than relying on a single chatbot, Full Shelf employs a fleet of agents, each with a well-defined role and limited tools. This separation ensures that each agent focuses on its specific task, preventing any agent from making unauthorized decisions. The system uses Google Cloud services, such as Vertex AI, Google ADK, Model Armor, Cloud Run, Cloud Spanner, Cloud KMS, Pub/Sub, Cloud Tasks, Cloud Scheduler, Cloud Trace, and Cloud Logging, to orchestrate the agents and maintain the integrity of the operating state. This architecture ensures that only authorized changes are made, and any errors or missing evidence are handled appropriately.\n\nFull Shelf's architecture emphasizes clear contracts, bounded responsibilities, deterministic sequencing, and thorough testing. This approach leads to a more robust and reliable system, preventing unauthorized actions and ensuring that directors are presented with accurate information about the changes, uncertainties, and decisions required. As the system evolves, it can be expanded to cover additional aspects of food bank operations, such as donation intake, warehouse management, volunteer availability, and coordination across networks.",
  "summary": "#AllThingsAgenticHackathon #GoogleCloud #Gemini #AgenticAI A recent lettuce recall made me wonder what happens when affected food is already moving through a local food bank. What I found was not a simple inventory problem. Food banks are living logistics networks. Inventory, refrigerated transportation, custody, partner agencies, safety rules, and commitments to the community all have to stay…",
  "key_points": [
    "Full Shelf is an AI solution for managing food bank operations",
    "Five specialized agents adapt to real-time changes in inventory and deliveries",
    "Human approval required before AI agents execute any actions"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}