The Compounding Unlock

I've been struggling to articulate something I've been experiencing for about six months now. Why AI feels fundamentally different from every other productivity breakthrough I've had in 20 years of building software.

It's not that things are faster. It's that there's a loop. And the loop keeps accelerating.

The Aha Moment

It started back in November when we rebuilt Archie from scratch—a story for another time. I built Martha (named after my middle child, the ironies are too many to go into here) as the factory that builds Archie. And something strange started happening.

The first week, Martha was flaky. Unreliable. I'd queue up work, go to bed, wake up to half-finished tasks and bugs.

The second week, things got more consistent. More tasks completed overnight. Fewer errors.

By week three, I was waking up to nine completed projects. Work I used to supervise step-by-step was just done.

But here's the part that took me a while to understand: It wasn't just getting incrementally better. Each week, the capacity didn't increase linearly. It multiplied.

Week one: Maybe 2-3 solid completions overnight. Week three: 9 projects done autonomously. Week eight: I stopped counting because the constraint wasn't Martha's capacity anymore. It was my ability to queue up enough work to keep it fed.

That's when I realised I wasn't experiencing productivity improvement. I was inside a compounding loop.

I was on a call with Ben Balbale a few weeks ago, an investor I've known since the Xero days. He'd just discovered what I'd been living with for months.

"It was a big step for me," he said, "when I went from using one agent to using five. Now I literally have seven different projects going simultaneously. Some of which I start at midnight before I go to bed."

I could hear it in his voice. That moment when you realise this isn't just "faster." This is structurally different.

He'd stumbled into the loop. And I finally had language for what I'd been trying to explain.

The Three Dimensions

What we're seeing, both in how we're building Archie and what it means for accounting firms, comes down to three dimensions that don't add together. They create a loop.

And each time through the loop, everything gets stronger.

1. Autonomous

Each agent owns its task end-to-end. You describe what you want, and it figures out the how. Reads the codebase, plans the approach, writes the code, runs the tests. No hand-holding.

For accounting, this means taking a monthly close process or a payroll workflow and having it execute completely. One instruction leads to a reliable outcome, every time.

2. Parallel

Multiple agents working simultaneously. One's building UI while another's wiring APIs while a third is writing database migrations. Not sequentially faster. Actually at the same time.

For accounting firms with 500 clients, this means Archie is working on all 500 overnight. Not one after another. All at once.

3. Asynchronous

No one's blocking on anyone else. Each agent gets its own context, its own timeline. Work happens 24/7 without coordination overhead or timezone constraints. Results converge when they're ready, not when some synchronisation point forces them to.

For accounting, this means the work that used to wait until Monday morning, or until Sarah got back from vacation, or until the senior manager had time to review it just happens.

The Compounding Loop

Now here's the part that I think is hard to grasp, and why I keep using the word "loop" instead of "improvement."

These three dimensions don't just make each other better. They create a cycle that compounds on itself.

Better autonomous execution → enables more parallel work → which runs asynchronously → generates more learning → makes autonomous execution even better → enables EVEN MORE parallel work.

And round and round it goes.

Let me break down the mechanics:

Autonomous gets smarter. Each task the agents complete teaches them how to handle the next one better. Edge cases get handled. Patterns get recognized. The reliability increases with every execution.

Parallel scales further. As agents get more reliable autonomously, you can safely run MORE of them in parallel. What started as 5 agents becomes 9, becomes 20, becomes 100. The constraint isn't the technology. It's how much work you can queue up.

Asynchronous compounds continuously. Because it's always running, the improvement cycle never stops. It's learning and getting better 24/7. While you sleep, the loop is running. While you're in meetings, the loop is running. The acceleration is constant.

I've been living inside this loop since November 2025. Each day Martha gets better. Each week I can queue up more work. Each month the reliability increases exponentially, not linearly.

A Month's Worth of Work Done In A Day.

Let me give you a concrete example of what this looks like in practice.

HaysMac, an accounting firm in the UK, was spending 600 hours a month processing payroll for 1,200 clients. Take payroll from desktop software, produce reports, lodge the general ledger in Xero, do the work papers, approve the ledger.

We got 600 hours down to about two.

Not two hours per client. Two hours total. For all 1,200 clients.

That's a month's worth of work done in a day.

And here's what makes it really interesting: This isn't the ceiling. This is the first loop. As Archie processes more payroll, it gets better at handling edge cases. As it gets better, we can run it across more clients in parallel. As it runs across more clients, it learns faster. The loop accelerates.

The cost to serve collapses. The quality improves because there's consistency and no fatigue. The speed increases by orders of magnitude. And every cycle through the loop makes the next cycle more powerful.

A New Operating Model

So when we think about the accounting industry, the compounding loop of autonomous, parallel, asynchronous AI fundamentally changes what's possible.

If what used to take days and people now takes hours and agents, then is the talent shortage an issue? Or would a firm’s operating model change in such a way that 50 people to serve 500 clients will no longer be the norm? The constraint won't be headcount or hours in the day. The economics of running a practice will look completely different in 24 months than they do today.

The structural advantages of this compounding loop are too significant to ignore.

The firms that understand this early, that lean into building their practices around this new architecture, are going to have an absurd advantage.

How Many Cycles?

Here's what I'm watching for now: How many cycles through the loop does it take before the acceleration becomes undeniable?

Have you felt this shift yet?

Have you experienced that moment where you realise you're not just getting 20% more productive, but you're inside a loop that keeps accelerating every day?

Because once you feel it, you can't “unfeel” it. And the industry is going to look very different once enough people are living inside this loop.

Build on.