A personal multi-agent assistant

A multi-agent system on WhatsApp: an orchestrator routes every message to specialist sub-agents, one of which runs a real agentic tool-use loop across email, calendar, tasks, and memory. Built for myself, in daily production use.

Personal/internal system. It operates over my own private data, so this is the architecture and capability only - never the personal content, which stays private.

Problem

I wanted an assistant that takes a one-line WhatsApp message and turns it into a concrete action across the tools I actually use - find a document in my inbox, read it, put something on my calendar, chase a task - reasoning over my own context rather than answering in general.

Why multi-agent

Fast routing, deep tool-using execution and long-form writing are not the same task, and they do not want the same model or the same prompt. So instead of one bloated prompt, the system is an orchestrator with a registry of specialists, each focused and independently improvable.

The agentic loop

The operations agent runs a genuine tool-use loop, not a fixed script: it reads email and attachments, looks up contacts, manages the calendar, keeps tasks, and writes to a private memory - deciding for itself which tool it needs next.

"Find the document attached to that email, read it, and draft a reply to the right person with it attached." It searches the inbox, reads the PDF, looks up the contact, drafts an accurate email with the attachment - and sends it once I confirm.

An hourly schedule turns it from reactive to proactive: a morning briefing, reminders that fire on time, and a weekly writing task. It surfaces conflicts and overdue items unprompted, with a recommendation attached.

Design principles

• Every message becomes an action - it carries the action out, then reports exactly what it did.
• An honest capability manifest. It knows what it cannot do and says so, offering the closest thing it can, rather than over-promising.
• Safe by default. Outbound messages to people are whitelisted; everything else becomes a draft for me to send.
• Cost-tiered. The cheap fast model routes; the stronger model only runs the deep work.

Outcome

A working multi-agent assistant in daily use: an orchestrator, tool-using specialists, stateful memory and a proactive engine.

What it demonstrates: multi-agent orchestration, a real tool-use loop over live tools, stateful context, proactivity, and honest capability design - the patterns the field is building on right now.