Capstone 4: Your Personal Ops Agent

Most people start the day the same way: scan an overflowing inbox, figure out what's actually urgent, check the calendar, and write the same kinds of replies they wrote yesterday. It's an hour of low-leverage work that an agent can do before you wake up, as long as you can trust it not to embarrass you.

In this final capstone you'll build that agent for yourself. Every weekday morning it reads new email, sorts it by what needs you, drafts replies in your voice, and writes a one-page briefing with your schedule. It never sends anything on its own. Sending happens in a separate review tool, with you at the keyboard, approving each message. The whole design is about getting most of the benefit while keeping the one irreversible action firmly in your hands.

What you'll build#

Two small programs and a schedule:

• run.ts, the scheduled job. It reads new email since its last run, triages each message with structured output, drafts replies for the ones that need one, and writes digests/2026-11-03.md (or whatever today is) with priorities, schedule, and inbox summary.
• review.ts, your approval desk. It walks through pending drafts and lets you send, edit, discard, or keep each one.
• A cron entry that runs the job at 7am on weekdays.

Email and calendar sit behind small interfaces with mock implementations, including a phishing email and an email containing a planted prompt injection. You can run everything with only an Anthropic API key, then connect your real accounts when you're ready.

This capstone draws on Automating Repetitive Busywork, Scheduled and Background Agents, Human-in-the-Loop Design, and The Agent Threat Model.