Agent Infrastructure

Running multiple AI agents for different people and different jobs in one household: a personal mobile cockpit, a coaching agent for a second user, automation agents, shared memory. Naive setups leak one user's context into another's session, or leave backups and credentials dangling.

A multi-agent architecture with explicit trust boundaries: separate agent profiles per user and per job, and a fully siloed containerized agent for the second user, with no network ports, no shared mounts, capped CPU and memory, and her own API key with a monthly spend cap.

Shared memory lives in one plain-text vault rather than a cloud sync, so there is a single source of truth and nothing to drift.

Underneath it all, a three-machine backup chain: nightly staging with restricted read-only SSH keys, hardlink snapshots with integrity checks, and a watchdog that alerts a phone within five minutes of a failed backup, dead host, or stale snapshot.

Every secret has a defined location and never enters shared notes or chat. Remote access is VPN-only; nothing is port-forwarded. One agent's privacy rules were tested against scripted fake-data probes before going live, and the probes found real leaks, which were then fixed and re-tested to a clean pass.

Infrastructure honesty: restore tests with database integrity checks, seven of seven privacy probes passing, spend caps verified in the provider console. A standalone handbook documents every device, address, port, and schedule so the system runs without any AI involved.

Six automated services, three backup chains, and an alerting watchdog run unattended. The whole cost stack is held to about thirty dollars a month. The measure of this project is that nothing in it needs me anymore.

A full restore drill onto clean hardware, and an off-site backup copy.