Boo: A Scheduler Daemon for AI Agents

The Goal

I wanted my AI agents to do things while I was not sitting in front of them. A morning brief before the workday, an inbox triage every 30 minutes, a reminder that fires once and then deletes itself. Plain cron can run a command, but it has no idea an agent exists: no notion of which agent or model to use, no capture of the response, no recovery when the laptop was asleep at 9am. Boo is the missing layer between a schedule and an agent.

What I Built

• A Rust daemon with three schedule types: recurring cron expressions, one-shot at times (including natural language like “tomorrow 9am”, parsed via the agent CLI with confirmation), and fixed every intervals.
• Agent-aware jobs: each job can target a specific agent and model, run in its own working directory, and enforce a timeout that kills the process group cleanly if it overruns.
• Heartbeat with missed-run recovery: the daemon tracks a heartbeat so it can tell the difference between “nothing was scheduled” and “the machine was asleep”, then catches up on runs it missed instead of silently dropping them.
• A full operator surface: desktop notifications, run history and statistics over 24h/7d/30d windows, live view of active runs with PID and elapsed time, resume of an interactive session to follow up on a scheduled run, and install/uninstall as an auto-start service.

What It Demonstrates

Boo is systems programming in service of agentic workflows: process supervision, signal handling, cron and interval math, persistence across reboots, and OS integration for notifications and auto-start, all in Rust. It has shipped through five tagged releases (v0.5.1 at last count) and is the scheduler I actually use to run my own proactive AI tasks. It is the concrete answer to a question I care about: what does it take to make an agent reliable enough to trust unattended?