Building a Platinum-Tier AI Employee: From File Watcher to Hybrid Cloud Agent
The Problem
Most AI agent tutorials stop at "call an API and return a response." But real-world autonomous agents need to handle failures, respect human oversight, operate across cloud and local environments, and maintain state across sessions. I wanted to build something that actually works in production — not just a demo.
The 4-Tier Evolution
My Personal AI Employee started as a simple Bronze-tier file watcher and evolved through four distinct tiers, each adding production-critical capabilities:
Bronze: File Watcher
The first version watched a directory for new files and processed them with Claude. Simple, but it proved the concept — an AI agent that monitors inputs and takes action without human prompting.
Silver: Multi-Channel Input
Silver added Gmail and WhatsApp monitoring. The agent could now watch multiple input channels, prioritize messages, and draft responses. This is where I learned that channel management is harder than AI inference.
Gold: MCP Server Integration
Gold introduced 4 custom MCP (Model Context Protocol) servers:
- fte-email — Gmail API integration with OAuth2, draft management, and thread awareness
- fte-social — WhatsApp Business API with message queuing and rate limiting
- fte-odoo — ERP integration for business process automation
- fte-documents — Document processing with OCR, classification, and filing
Each server implements circuit breaker patterns for reliability and exponential backoff for rate limiting.
Platinum: Hybrid Cloud-Local Architecture
The Platinum tier is where everything came together. The architecture splits across cloud and local:
Cloud VM (Oracle Cloud): Runs the agent daemon, monitors all input channels, drafts responses, and stages actions. This runs 24/7 even when I'm offline.
Local Machine: When I reconnect, git-sync pulls the latest state. I review staged actions, approve or reject them, and the agent executes approved actions through the MCP servers.
This hybrid approach solves two critical problems:
- Always-on monitoring without running my local machine 24/7
- Human-in-the-loop safety for all consequential actions
Engineering Patterns That Matter
Circuit Breakers
Every MCP server connection uses a circuit breaker with three states: closed (normal), open (failing, stop trying), and half-open (test one request). This prevents cascade failures when a service goes down.
Correlation IDs
Every action chain gets a unique correlation ID that flows through logs, drafts, and approvals. When I'm reviewing 20 staged actions, I can trace each one back to its trigger.
Human-in-the-Loop (HITL) Safety Levels
Actions are classified into three safety levels:
- Routine — Auto-executed (filing documents, tagging emails)
- Sensitive — Staged for review (sending emails, updating records)
- Critical — Requires explicit approval with confirmation (financial actions, deletions)
PM2 Daemon Management
The cloud agent runs under PM2 with automatic restart, log rotation, and health monitoring. If the agent crashes, PM2 brings it back up within seconds.
What 21 ADRs Taught Me
I wrote 21 Architecture Decision Records (ADRs) for this project. Each one documents:
- Context — What problem we're solving
- Decision — What we chose and why
- Consequences — What trade-offs we accepted
The most valuable lesson: ADRs force you to think before you build. Every time I was tempted to "just code it," writing the ADR first revealed edge cases I hadn't considered.
Key ADRs that shaped the architecture:
- ADR-003: Hybrid Cloud-Local Split — Why we can't just run everything locally or everything in the cloud
- ADR-007: Circuit Breaker Implementation — Why we chose the half-open state pattern over simple retry
- ADR-012: HITL Safety Levels — How we classify actions and why three levels is the right granularity
- ADR-019: Git-Sync State Transfer — Why git is the right protocol for cloud-to-local state transfer
The Verified Live Demo
The hackathon judges didn't just review code — they watched the system work:
- Sent a test email to the monitored Gmail account
- Watched the cloud VM detect and process it within 30 seconds
- Triggered git-sync on local machine
- Reviewed the staged draft response
- Approved the action
- Confirmed the response was sent via the MCP email server
This end-to-end demonstration earned the Platinum tier — the highest achievable rating.
Key Takeaways
- Start simple, evolve deliberately. Each tier solved real problems from the previous tier.
- Production AI needs safety rails. HITL gates aren't optional — they're essential.
- Hybrid architectures solve real constraints. Cloud for availability, local for control.
- Document decisions, not just code. ADRs are the most underrated engineering practice.
- Circuit breakers save production systems. External services fail. Plan for it.
Related Projects
Check out the full project details at Personal AI Employee or explore the source code on GitHub.