Skip to content
AZ

ai-employee-autonomous-systemPython857 KBDONE

Personal AI Employee (Digital FTE)

An autonomous digital employee. Watches Gmail, WhatsApp, the filesystem and finances, plans its own work, asks permission before it acts — and never silently drops a task.

Four watchers monitor Gmail, WhatsApp, the filesystem and finances. Anything they notice enters the vault as a markdown task file, and an orchestrator runs a READ → PLAN → WRITE → EXECUTE reasoning cycle over it. Actions execute through a custom MCP layer of three hand-written servers — Gmail, WhatsApp and LinkedIn.

Human-in-the-loop by default

Payments, new contacts and social posts always stop at an approval gate. The system is autonomous about deciding what to do and deliberately not autonomous about doing the things that are expensive to undo.

The Ralph Wiggum loop

A failed task is not dropped and not retried blindly. It is re-injected into the planner with its own error as context, up to ten times, and escalated to a human after that. The failure becomes part of the next prompt, which is why the retry is worth anything.

The vault is the state machine

Task files move between folders, and the folder is the status:

Inbox → Plans → Pending_Approval → Approved → Done

with Needs_Action for retries, and Rejected / Review_Required as terminal states. Nothing about the state lives in memory — it survives a restart because it is a file.

Operations

Every action appends to audit.jsonl. A watchdog restarts crashed processes. An Obsidian dashboard updates itself from the vault. With no AI key present the system still runs, falling back to heuristic plans rather than refusing to start.

highlights

  • Human-in-the-loop approval gates on payments, new contacts and social posts
  • Ralph Wiggum loop — failed tasks retried with error context, 10× then escalated
  • Vault task flow: Inbox → Plans → Pending_Approval → Approved → Done
  • Every action written to audit.jsonl, with a watchdog restarting crashed processes
  • Runs without an AI key via heuristic fallback plans

built with

  • Python
  • MCP
  • Gmail API
  • WhatsApp
  • LinkedIn
  • Obsidian
All work