My daily driver: a self-hosted AI-agent environment running ten isolated profiles on Arch Linux — one per job (tutoring, video production, career, coursework), each with its own memory, skills, and cron jobs.

Requests route across multiple LLM providers with cost-based fallbacks, so a cheap flash model handles routine work and a frontier model only spins up when it earns it. Parallel research subagents fan out with citations, and reusable procedures are packaged as tested Python CLI tools shared across profiles through MCP integrations.

The interesting problems were operational, not glamorous: memory that survives restarts without leaking between profiles, a watchdog for the memory store, and provider failover that doesn't silently downgrade quality.