The phrase "AI agents" has been hyped to the point of meaninglessness. Half the time it means a chatbot wrapper. The other half it means an enterprise vision pitch about "the agentic future" that's still 18 months away.
I'm going to skip both and tell you about the seven AI agents I have running on my own machines and accounts right now. Four are personal. Three are business. None are theoretical. Each handles a job I used to do manually that I never want to do again.
The shared trait: every one started as a checklist. If I could describe the workflow in five bullet points, I could turn it into an agent.
1. The "kid pickup logic" agent (personal)
The job it does: Watches my Outlook calendar for the day, cross-references my wife's calendar, the kids' school schedule, and my Cal.com bookings, and surfaces conflicts the night before. Specifically: who's picking up which kid from which activity, and does anyone need to leave a meeting early.
Stack: n8n + Microsoft Graph API + a small Claude reasoning step. Runs at 9pm. Drops a 4-bullet summary into our shared Slack DM.
Cost: ~$0.04 of API tokens per night.
Where it breaks: When the school posts schedule changes only to the parent portal and not the calendar feed. Solved with a manual "remind me about anything new from the school" question every Monday.
This is the most-used automation in my life and the one nobody else knows I have. It's saved more arguments than it has any right to.
2. The photo library tagger (personal)
The job it does: Looks at every photo I import to my Apple Photos library, runs it through a vision model, and writes structured tags into the photo metadata — locations, people, activities, document scans, screenshots vs. real photos.
Stack: A LaunchAgent on macOS watching ~/Pictures/Photos Library.photoslibrary/. Sends new images to Claude Sonnet. Writes tags via exiftool.
Cost: ~$3-5/month at my photo volume.
Where it breaks: Burst photos eat tokens fast. I throttle by only processing every 5th photo in a burst, then doing the rest if I open the burst manually.
The unlock isn't the tags themselves — it's the searches I can finally do. "Show me all the photos of the kids' artwork from 2024" now actually works.
3. The inbox triage + draft-reply agent (personal/business)
The job it does: Sorts incoming email into Reply / Read / Archive / Investor-grade-stuff. For the Reply pile, drafts a one-paragraph reply in my voice, saved to Gmail's draft folder for me to review.
Stack: Gmail API + a custom n8n flow + Claude. Runs every 30 minutes during work hours.
Cost: ~$2/month of API.
Where it breaks: Cold sales pitches sometimes get past the filter and land in Reply. The agent then drafts polite-but-pointless responses I'd never send. Acceptable — I delete them in 3 seconds, beats missing a real one.
The rule I follow: the agent never sends, only drafts. I review every reply before it leaves. The 30 seconds per email is worth it for the 5+ minutes per email it saves on initial composition.
4. The monthly budget categorization agent (personal)
The job it does: Pulls my bank and credit card transactions from last month, categorizes them, flags anomalies (subscriptions I forgot about, unusual recurring charges, "is that really $80 at the same coffee shop in one week?"), and writes a one-page monthly summary.
Stack: Plaid API for transaction pulls + Claude for categorization + a simple Notion page for output.
Cost: Plaid dev tier (free for personal use) + ~$1/month API.
Where it breaks: Some ambiguous merchants (PAYPAL*SUMTHIN) get miscategorized. I correct manually once and the agent learns the mapping for next month.
It's an off-the-shelf job — Mint and YNAB do something similar — but the LLM-generated summary is what makes me actually read it. "You spent 18% more on dining than last month, mostly from three Friday-night events" is more useful than a pie chart.
5. The cold-call follow-up agent (business — DealerScout)
The job it does: After DealerScout.ai's outbound voice agent finishes a call, this agent reads the transcript, writes a structured CRM update, drafts a personalized follow-up SMS or email, and schedules the next touchpoint based on the call outcome.
Stack: Retell AI webhook → n8n → GPT-4.1 → dealer's CRM API (varies — VinSolutions, DealerSocket, etc.).
Cost: ~$0.15 per call processed. Built into DealerScout's per-seat pricing for clients.
Where it breaks: When the dealer's CRM API is down (more often than you'd think — looking at you, certain legacy DMS vendors). We queue and retry; nothing is ever lost.
This agent is the unsexy half of voice AI. The cold call gets all the press. The follow-up is what actually books the appointment.
6. The "research → markdown brief" agent (business — consulting)
The job it does: I hand it a topic ("competitive landscape for AI scheduling tools," "what's new in Vercel pricing for hobby projects"), it does 6-12 web searches, reads the most credible sources, and gives me a 1-2 page markdown brief with citations.
Stack: Claude Code with web tools enabled. I literally type the request and walk away for 5 minutes.
Cost: Bundled into my Claude Pro subscription.
Where it breaks: Topics where the sources are mostly paywalled (industry reports, gated whitepapers). I supplement with what I can pull from my own subscriptions.
This is the agent that's most quietly changed my consulting workflow. What used to be a 90-minute "let me research that and get back to you" cycle is now a 10-minute prep + the actual call.
7. The code review agent (business — solo SWE work)
The job it does: Before any commit on my own projects, runs a code review pass — checks for security gotchas, simplification opportunities, and obvious test gaps. Posts comments inline.
Stack: Claude Code's /security-review and /simplify skills, plus a custom git pre-commit hook.
Cost: Bundled into my Claude Pro.
Where it breaks: Long-running sessions occasionally hit context limits and miss the bigger picture. I run it on smaller, scoped diffs to avoid this.
I'm a 20-year SWE. I don't need an AI to write my code. I do appreciate having a second set of eyes on every commit, especially when I'm shipping solo.
The rule of thumb
If you can describe a workflow as a checklist with fewer than 10 steps, you can probably turn it into an agent.
If the steps require live judgment ("decide whether this client needs a $15k engagement or a $3k one"), keep yourself in the loop and let the agent surface the data, not the decision.
The boundary between "useful agent" and "creepy / dangerous agent" is exactly where you let the system act vs. where you let it suggest. Get that line right and AI becomes the most reliable junior teammate you've ever worked with. Get it wrong and you'll spend more time cleaning up after it than it ever saved.
Have a workflow in mind that's eating your week? I help individuals, founders, and SMB teams design AI agents that actually fit their reality — personal or business. Book a 30-min discovery call and we'll find your highest-leverage agent in 30 minutes flat.