My AI Agents Now Improve Themselves. Here's How I Built Self-Expanding Flywheels.

What's a Flywheel (And Why Most Founders Get It Wrong)

Every founder I coach says they have flywheels.

I've invested in 230+ startups. I've coached hundreds of founders. And I hear the same thing constantly: "We have a flywheel for content." "Our product has a built-in growth flywheel."

No, you don't.

You have a process. You have a workflow. Maybe you have a good habit.

But a flywheel is something specific: it compounds without your input.

Most "flywheels" are just recurring tasks someone does manually. That's not a flywheel. That's a to-do list.

A real flywheel has three properties:

1. It runs automatically (no human trigger needed)
2. It measures its own output (knows what worked and what didn't)
3. It adjusts based on the measurement (gets better over time)

If your "flywheel" needs you to push it every time, it's a hamster wheel.

The Blog Writer: My Best Flywheel

The best example I've built is my blog writer agent.

Here's how it works:

Step 1: It pulls from the Batko Brain. Every piece of content I've ever written, 4.1 million words across 57 tables in a SQLite database. It knows my voice, my opinions, my examples.

Step 2: It gets topics from my coaching sessions. When I coach founders, patterns emerge. The same questions come up. The same mistakes. The same breakthroughs. Those patterns become blog topics automatically.

Step 3: It checks what's working. It looks at my blog analytics. Which posts got views. Which got clicks. Which got shared. Which ones people actually read to the end.

Step 4: It writes more of what works. Not copying. Not repeating. But understanding what resonates and leaning into those themes, those formats, those types of arguments.

Then it loops back to Step 1 and starts again.

I don't tell it what to write. I don't review its topic list. I don't manually check analytics.

It closes its own loop.

That's a flywheel.

Why "Close the Loop" Changed Everything

Before that moment with Claude, every agent I built was one-directional.

Input goes in. Output comes out. Done.

My Monday standup agent would brief me every week. Same format. Same depth. Whether I read it or not.

My email triage would sort my inbox. Same rules. Same categories. Whether the categories still made sense or not.

They worked. But they didn't improve.

The moment I added a feedback loop to each agent, everything changed.

Now my Monday standup tracks which sections I actually act on. If I consistently ignore the "Design" section but always act on "Revenue", the agent adjusts. More revenue analysis. Less design fluff.

My email triage notices when I override its categorisation. If I keep moving emails from "Low Priority" to "Urgent", it recalibrates.

Every agent now loops back and checks its own performance.

They don't just run. They learn. Not in some abstract machine-learning sense. In a practical, "did the human actually use this output?" sense.

The Cost Flywheel Nobody Talks About

Here's one most people miss: cost optimisation as a flywheel.

I get a weekly email reviewing my AI costs. Token usage across every agent. Vercel hosting costs. API calls. Storage.

But it's not just a report. It's a flywheel.

The email identifies which agents are burning tokens inefficiently. It suggests optimisations. It tracks whether last week's optimisations actually reduced costs.

When I restructured the Batko Brain from one massive context load to FTS5 search queries, token costs dropped 80%. The cost flywheel spotted the problem, suggested the fix, and verified the result.

Now it does that continuously. Without me asking.

Build a flywheel that will never fail.

That's not just a nice quote. It's a design principle. Every agent I build now has a cost check built in. Every workflow has a "was this worth the tokens?" loop.

Because automation that gets more expensive over time isn't automation. It's a liability.

The Flywheels I Haven't Built Yet

I'm being honest here. I have a list of flywheels that should exist but don't yet:

Email drafting flywheel. After every meeting, draft a follow-up email based on the transcript. Check whether I actually sent it. If I edited it heavily, learn from the edits. If I sent it as-is, note that the draft was good enough.

Client pitch flywheel. When I meet a potential AI OS client, automatically research their business, identify automation opportunities, and draft a personalised pitch. Track which pitches converted and adjust the template.

Inbox review flywheel. Not just sorting email. Tracking response patterns. Which emails do I always reply to within an hour? Which ones sit for days? Automatically prioritise based on my actual behaviour, not my stated preferences.

Each of these follows the same pattern: automate the action, measure the result, adjust the approach.

None of them are built yet. But the pattern is clear. And the pattern is the point.

How to Build Your First Flywheel

If you're building AI agents that just run but don't improve, here's how to add the loop:

1. Start with what you already automate

Pick an agent or workflow that's already running. Don't build from scratch. Add a feedback mechanism to something that works.

2. Define "good output"

For my blog writer, good output = posts that get read to the end. For my email triage, good output = categorisations I don't override. For my cost flywheel, good output = lower token spend with same results.

You need a measurable signal. Not a feeling.

3. Log the signal automatically

Don't rely on yourself to manually rate outputs. Build the measurement into the workflow. Page views are automatic. Email open rates are automatic. "Did the human edit this?" is automatic.

4. Feed it back

This is where most people stop. They measure but don't close the loop.

The signal needs to flow back into the next run. Your agent needs to say: "Last week's output scored X. Here's what I'll change this week."

5. Set it and check monthly

A good flywheel doesn't need daily attention. Set it up. Let it run. Check in once a month to make sure the feedback loop is actually improving things.

If it's not improving, the signal is wrong. Change what you're measuring.

The Bigger Picture: Compounding Leverage

I run a one-person company. Every flywheel I build is one less thing I need to think about.

Not just one less thing I need to do. One less thing I need to think about.

That's the difference between automation and flywheels.

Automation saves you time. Flywheels save you attention.

And attention is the only resource that actually matters when you're running everything yourself.

Every flywheel I build frees up cognitive space for the work only I can do: coaching founders, making investment decisions, creating content that comes from real experience.

The agents handle everything else. And they get better at it every week.

Leverage beats labour. Every time.

Six months ago I was doing all of this manually. Now I have agents that improve themselves while I sleep.

That's not magic. It's just closing the loop.