By clicking “Accept”, you agree to the storing of cookies on your device to enhance site navigation, analyze site usage, and assist in our marketing efforts. View our Privacy Policy for more information.

Systems

Start With the Brain, Not the Agent

AI agents are overhyped for most business owners. Here's how I run my business on an AI second brain with scheduled runs, and how to start yours in an afternoon.

Joshua McSorley5 min readNovember 2026

Everyone is building AI agents right now. Most of the business owners I talk to don't need one yet.

Here's the pattern I keep seeing. An owner gets sold on autonomous agents. They spend weeks wiring up tools, prompts, and guardrails. They end up with something fragile that nobody trusts enough to run in production. Or worse, they do run it, and it quietly erodes their customers' trust.

Meanwhile, the thing that would actually help them is a lot less glamorous. It's a second brain: one place where everything about the business lives, with AI reading and writing to it every day. I've been running my own business this way for months, and it does more for me than any agent I've seen.

This is how it's set up, what broke along the way, and how you can start one in an afternoon.

The real problem isn't automation

When I sit down with service business owners, the bottleneck is almost never "we need something to act on its own." It's that nobody can find anything.

Projects live in one tool. Client notes live in another. Follow-ups live in someone's inbox, decisions live in a meeting nobody wrote down, and the rest lives in the owner's head. Every week, they lose hours rebuilding context they already had.

An agent doesn't fix that. An agent with no context just makes confident mistakes faster. What fixes it is giving AI something to remember.

What I built

My second brain is an Obsidian vault. It's a folder of plain markdown files, nothing more. That matters: any AI tool can read it, and nothing is locked inside one app's database.

Everything I work on lives there, personal and professional. Projects, people, companies, meetings, strategy, pricing, billing, even the book I'm writing. If the vault and any other tool disagree, the vault wins.

A few files do most of the heavy lifting:

  • AGENTS.md is the rulebook. It tells the AI how every folder works and where things go. Every run reads it first.
  • TASKS.md is my only task list. Every task from every source ends up here.
  • BILLING.md tracks invoices, retainers, and billable work.
  • daily/ holds one note per day. Everything the AI produces lands there for me to review.
  • raw/ is an inbox for transcripts and files the AI hasn't processed yet.

The vault sits in a private GitHub repo, and Obsidian on my laptop syncs to it. That one decision turned out to be more important than I expected, but more on that below.

How it stays current without me

A second brain is only useful if it's up to date, and I'm not going to update it by hand. So the AI pulls from the tools I already use: Gmail, Google Calendar, Google Meet (with Gemini notes), Granola for meeting transcripts, my CRM, and Stripe for billing. I use Claude, with a saved skill that knows how to add tasks, process meetings, log billable work, and update projects from any chat.

Then a handful of scheduled runs do the routine work:

RunWhenWhat it does
Morning briefWeekday morningsReads email, tasks, and calendar, then tells me what matters today.
End-of-day captureEvery eveningPulls tasks, decisions, and billable work out of the day's meetings and email.
Weekly planMondaysLays out the week and lists who I owe a follow-up.
LinkedIn connectionsMondaysChecks new connections against the vault and drafts a message for each.
Monthly reviewOnce a monthSteps back on projects, billing, and what's slipping.

Follow-ups run on a rhythm by relationship: prospects every two weeks, partners monthly, peers every two months, and past clients quarterly. I don't have to remember any of it. It just shows up in my daily note.

That's the whole system. No orchestration layer, no agent swarm. I call it agentic light: AI does the remembering, sorting, and drafting, and I make the calls.

The rules that make it trustworthy

The reason I let this run every day is that it can't embarrass me. A few rules make sure of that.

Nothing goes out without my approval. The AI drafts emails, messages, and follow-ups into my daily note. It never sends, publishes, or deletes anything on its own. That single rule is the difference between a helpful system and a liability.

The rulebook gets fixed, not just the note. When the AI files something in the wrong place, I update AGENTS.md so it doesn't happen again. That's how the system gets smarter every week.

One of everything. One task list, one source of truth, one place to look. The moment you have two, you have zero.

What broke along the way

This didn't work perfectly out of the gate, and I think that's worth saying.

My scheduled runs used to depend on my laptop. If it was asleep or closed, runs got missed, and the vault quietly fell behind. Moving the vault to a GitHub repo, so the runs could happen in the cloud, fixed that. I also added a watchdog that re-runs any missed job and notifies me, because scheduled jobs fail silently.

I also had too many places holding information. I retired Notion completely so the vault could be the only source of truth. And I learned to capture before I automate. Meetings turned out to be the richest input by far, so processing transcripts well paid off before any fancy automation did.

How to start yours

You don't need my whole setup on day one. Here's the afternoon version:

  1. Make the folder. Install Obsidian and create a vault with a TASKS.md file and a few folders: projects, people, meetings, and daily.
  2. Write your AGENTS.md. One page in plain English explaining what goes where. This is the most important file you'll write.
  3. Connect an AI that can read and write files. Give it access to the vault, your email, and your calendar.
  4. Feed it one week of meetings. Have it pull out tasks, decisions, and people. You'll see the value fast.
  5. Add one scheduled run. Start with an end-of-day capture. Add a morning brief once you trust it.

Then let it compound. Every week the context gets richer, and every week the AI gets more useful. That's the part most agents never reach. They're stuck starting from zero every time, because they have nothing to remember.

So if you're stuck somewhere between AI hype and AI that actually helps, start with the brain, not the agent. The agents can come later, and when they do, they'll have something to work with.

JM
Joshua McSorley

Marketing Systems Consultant. I help service businesses find and fix the gaps between their website, leads, and sales.

Ready to dig in?

Most service businesses have three or four of these gaps at once.

Run the free website assessment to see which ones are costing you leads right now. Takes about 60 seconds. No email required to see your results.

Run the assessment