Inspiration
The rapid ascent of autonomous AI agents promises unprecedented convenience, but until now, multi-agent deployments have been confined to expensive cloud architectures, heavy databases, and recurring per-token API bills. For neighborhood small businesses—cafés, bakeries, indie bookshops—these costs create an insurmountable barrier.
We asked: What if a neighborhood coffee shop could deploy a collaborative 5-agent network with zero server hosting costs, complete customer privacy, and native machine discoverability?
Built for the Cognizant Open Agentic Web Challenge, Bean & Leaf Coffee brings this vision to life. By leveraging standard web primitives (HTML, CSS, JavaScript, JSON, and HTTPS), Bean & Leaf provides a dual-interface platform serving human visitors while allowing external personal agents (like an autonomous morning routine assistant or delivery aggregator) to discover, query, and transact across origins.
What It Does
Bean & Leaf Coffee operates as an autonomous, browser-executed agent network:
The 5-Agent Collaborative Network:
- ☕ Coffee Concierge (
front_man): The primary greeter and intelligent coordinator. Understands customer intent, answers general questions, and seamlessly routes to specialized sub-agents. - 🕐 Hours Agent: Computes real-time store open/closed status against a weekly operating schedule and holiday calendar.
- 📖 Menu Agent: Catalogs artisanal coffees, teas, and bakery items with certified dietary filters (100% vegan, gluten-free, and dairy-free).
- 🧾 Order Agent: Validates selections, calculates item totals, generates a unique pickup ticket (
#1000–#9999) with estimated prep time, and publishes it to the kitchen queue. - 🪑 Reservation Agent: Checks seating capacity across 24 indoor and patio seats and confirms reservations.
- ☕ Coffee Concierge (
Zero-Host-Cost BYOK Runtime (
runtime.html):- The café pays $0 for AI inference. Visitors supply their own API keys (Google Gemini, OpenAI, Anthropic, or local Ollama), or select our instant 1-click Mock engine (no key required).
- Execution is sandboxed in the customer's browser; credentials and chats never touch an intermediary server.
Machine Discoverability (
agent-manifest.json):- Standardized manifest linked via
<link rel="agent-manifest">allowing web crawlers and autonomous user agents (Siri, Otari, personal shopping bots) to discover the café's capabilities.
- Standardized manifest linked via
Live Public State Board (
orders.json):- An open kitchen display system (KDS) where human customer orders and external agent orders converge in real time.
Cross-Site Agent Interoperability Simulator:
- A built-in testing suite demonstrating external agent handshakes (e.g. an autonomous morning routine bot discovering the manifest, negotiating an order, and receiving a pickup ticket programmatically).
How We Built It
- Conforming Architecture: Built to the Cognizant AI Lab Worldwide Agentic Web MVP Specification (
worldwide_agentic_web_mvp_spec.md). - Declarative Manifests: Pure data-driven JSON (
agent-manifest.jsonandagent-network.json) declaring roles, system prompts, capabilities, and store schemas. - Client-Side Sandbox: Vanilla HTML5, modern CSS, and asynchronous JavaScript with zero external runtime dependencies.
- Multi-Model BYOK Pipeline: Direct browser REST implementations for Google Gemini (
gemini-2.0-flash), OpenAI (gpt-4o-mini), Anthropic (claude-3-5-haikuwith direct-browser access headers), and local Ollama (llama3.2). - State Synchronization: Dual-tier storage syncing initial seeded
orders.jsonstate with local browser storage and live UI updates.
Challenges We Ran Into
- Browser Sandbox & CORS: Direct-browser BYOK requires navigating strict cross-origin restrictions, such as Anthropic's mandatory direct-browser headers and Ollama's local origin policies.
- Decentralized State on Static Hosting: Static hosts like GitHub Pages serve read-only files. We designed an interoperable browser simulation layer that persists newly placed orders in local storage while keeping the public
orders.jsoninterface consistent. - Intent Routing Without Server Orchestrators: Enabling the front-man Concierge to route user queries gracefully across both raw LLM outputs and mock rule engines.
Accomplishments That We're Proud Of
- True Zero Operating Cost: Complete 5-agent business functionality running on static web infrastructure with zero ongoing cloud bills.
- Privacy by Design: API keys and prompts remain in browser memory; neither the shop nor any third party intercepts user data.
- Real Cross-Site Interoperability: Proving that external agents can discover and transact with small businesses over open standards without proprietary API silos.
What We Learned
The Open Agentic Web demonstrates that AI does not need to be locked behind centralized mega-platforms. By using open JSON manifests and browser-level execution, the web itself becomes the coordination layer for autonomous agents.
What's Next for Bean & Leaf Coffee
- Model Context Protocol (MCP): Exposing the menu and table inventory as standard MCP tool endpoints.
- Autonomous Micro-Payments: Adding Web Monetization or Lightning settlement so external agents can pay autonomously during ticket generation.
- Federated Local Networks: Linking Bean & Leaf to neighboring local businesses for cross-shop itineraries (e.g. "coffee + bookstore").
Built With
- anthropic
- byok
- css3
- gemini-api
- github
- html5
- javascript
- json
- multi-agent
- ollama
- open-agentic-web
- openai