How to Build a Self-Improving Business

A self-improving business runs on AI loops: agents that are given a goal, iterate towards it on a schedule, and learn from their own results. This essay explains what a loop and a harness are, and gives you a five-step framework (Read, Regulate, React, Review, Refine) for building one around any goal in your company.

Loops are one of the most interesting AI concepts to emerge over the last few months.

I learned about loops after playing with Ploy.ai, a new startup which rebuilds your website and then optimises it every night, based on analytics and Search Console data.

It can fix technical issues, create new SEO landing pages, or set up A/B tests.

I tried to recreate this using Claude Cowork, but a few hours later, I realised that it was more complicated than I’d realised.

Building a daily optimisation loop takes time to get right.

Plus, a company like Ploy isn’t just running one loop. It’s running a whole series of different loops, each optimising its own little part of the marketing stack.

I gave up on rebuilding Ploy in-house, and I’m focusing my attention on creating loops for my core business: making CEOs formidable.

In any case, the idea of AI working through the night to optimise and iterate towards a measurable goal is compelling.

So how should you think about loops?

Loops and Harnesses

A loop is where you give an AI agent a goal and get it to iterate towards achieving it.

For example, you could give it a goal to improve a landing page’s conversion rate. The agent could collect data about the page, make intelligent changes and measure the result to see whether they worked.

It can then repeat the process on a regular schedule, say, every night. That’s the loop.

You can create a simple loop by putting an agent on a schedule in your AI tool of choice.

However, if you want to create more sophisticated loops, you’ll need to build your own harness.

A harness is a combination of AI and deterministic code that controls and supports how the AI works.

Technically, Claude Cowork is itself an agent harness, and scheduling is one of its features.

But harnesses can do more than trigger agents.

They can manage state, streamline access to data, enforce permissions and sequencing, automatically stop or roll back a process, or escalate to a human when something goes wrong.

To get your head around designing loops and harnesses, I came up with a five-step framework you can use to architect a loop around one of your goals.

1. Read

One of the defining features of an AI company is that all data is readable by AI.

This includes the results of an agent’s work.

Let’s take a customer-service agent as an example.

You’ll want the agent to read customer data, communication channels, company policies and perhaps even the source code to verify the product’s live capabilities.

But to measure its own performance, it’ll also need to collect customer-satisfaction scores or track the time taken to resolve a ticket.

Question: How can you measure the end-result you want to optimise?

2. Regulate

Regulation is about the hard constraints applied before or during an action.

Since a loop is meant to run autonomously, you’ll want to define the rules it must abide by, the sequence of steps you want it to follow, and what it needs to log.

Imagine a customer-service agent operating in a regulated industry. There might be strict data-protection policies that need to be followed.

The AI’s harness can enforce these policies using deterministic code that checks messages for sensitive customer data. It might also use trained evaluation agents that assess the output for compliance.

Question: What are the constraints you want the AI agent to respect?

3. React

With both data and rules, the AI can now react to the situation.

To do this, it may need access to tools that allow it to send emails, write code or access other systems.

If the system doesn’t have an MCP or CLI interface, the harness might make it possible to work directly with an API.

For example, a customer-service agent might respond to a customer query and update internal systems to resolve the issue.

It might even update the code to fix the underlying customer issue, depending on the constraints you’ve set.

Question: What tools does AI need to do a great job?

Get my weekly CEO essay ->

4. Review

How do you ensure quality? This is where the review step comes in.

Reviews might be performed by a human, but they could also be done by a trained evaluation model that knows what to look for.

For our customer-service agent, an evaluation model could score conversations against a customer-service rubric and identify recurring weaknesses.

This could be combined with human spot checks and escalation whenever bad customer feedback is received.

Question: Which parts of the process require a quality check?

5. Refine

Once the results are in, the loop can work out how to improve its approach. You can think of this as the experimentation step.

This happens on a schedule, often overnight so changes are ready when the team wakes up.

In the case of a customer-service loop, the AI could compare all the positive and negative scores from a given week and develop a testable hypothesis for improving the following week’s scores.

An underrated source of input data is the conversations between humans and AIs in the system. This can often reveal additional opportunities for improvement.

Why Loops Matter

Recently, I brought together some CEOs from my network to discuss how they’re approaching AI in their companies.

Right now, a lot of companies are in the workflow-mapping stage.

This is where you find your best people and map what they do to achieve their goals.

Loops are subtly different. Rather than automating what people do today, they are designed to iterate by themselves towards achieving a goal.

A well-designed loop can iterate faster than humans and, set up well, a loop could eventually outperform a human.

And if you string enough of them together, you begin to create a self-improving business.

For now, at least, humans are needed to ensure agents are pursuing the right goals and in the right ways.

However, for how long this remains true is anyone’s guess.

I’d love to learn about AI loops you’ve created in your business and whether you’ve seen any surprising results.

Continue Learning About AI in Your Company

• Want to see loops in practice? Read why I built an AI organisation.
• Struggling to get your team using AI? Try these 10 new ways to drive AI adoption.
• Want a simple place to start? Learn how smart CEOs analyse their meetings.

Join my $100M CEO newsletter ->

Originally published August 27, 2026.