Always-On Agent System

The Problem

Managing multiple AI tools across different platforms is fragmented. You have one chatbot here, another automation there, and none of them talk to each other. There's no central intelligence that understands the full picture and can coordinate across tools and channels.

The Approach

Hub-and-spoke on a Mac Mini M4. One coordinator agent has the full picture and delegates to specialists: one for coding, one for research, one for analysis. I pick the model per task based on what it actually needs. I use this system every day across Discord, Telegram, WhatsApp, and iMessage.

Key Design Decisions

1. 1.Model routing: expensive reasoning models for complex analysis, fast cheap models for routine lookups. The difference between a $0.50 task and a $5 task is usually which model you pick, not how you prompt it.
2. 2.Persistent memory: structured markdown files that survive session resets. Context carries across days and weeks. The agents write to these files, but the schema is mine.
3. 3.Monitoring: security scans, system checks, and alerts run on a schedule. I set the cadence and escalation rules. Most things get ignored quietly. I only hear about actual problems.
4. 4.Tool integration: GitHub, 1Password, search APIs, browser automation, PDF analysis. Each tool is there because my workflow actually needs it, not because it looked cool in a demo.

Results

Stack