Progress Update on Making an Automated AI Business

TL;DR; First two ‘agents’ are mostly complete, through a mixture of blended AI tool use and hard-code. They still need human input, and so I am now staggering my approach to allow this, then to later replace myself with AI.

What’s clear is that this is going to take a much more organised setup than I’ve been using so far. So I will continue to:

• Try lots of approaches and tools
• See what sticks
• Report it all to you

AI Directory Maker Update

Scrappy Blend of AI Tools

I’ve already tested lots of AI tools for this challenge (e.g. Gumloop, Relevance AI, API’s etc.) and it’s clear that by the end of the year this challenge will be easier to pull off.

For now though, I’ve adopted a scrappy AI tool blend as the way forward.

It seems no single tool will be able to pull off a fully automated business, and so it’s going to take an orchestra of different tools – the challenge being to make them play in harmony.

Here’s what the first two agents look like now:

AI Directory Prospector Agent

AI Directory Prospector: Step	AI Directory Prospector: Status
Start: Capture initial input: user describes niche & picks best target idea.	Working
Explore Trends: Analyze current trends for the chosen niche or idea list.	Working
Niche Research: Gather deeper niche data and validate potential opportunities (nichereport.io)	Deferred
Search Engine Research: Use GPT or external APIs to explore search volume, competition, and related keywords.	Working
Make Decision: Decide whether to proceed or abort based on research data.	Working
Domain Prospecting: Check domain availability or auctions for relevant domain names.	Working
Final Output: Provide the final result with chosen domain, idea, keyword data.	Working
Fail Output: Handles any errors or dead ends by reporting failure.	Deferred

AI Directory Builder Agent

AI Directory Builder: Step	AI Directory Builder: Status
Retrieve Starting Data: Gather inputs from the Prospector Agent (domain choice, niche info, etc.).	Working – needs human sign-off
Register Domain: Register the chosen domain via DNSimple & point to hosting.	Working
Set Up Hosting, SSL, and Email: Provision hosting, configure SSL certificates, and create email accounts.	WIP
Branding: Generate or apply branding assets (logo, color palette, style guidelines).	Working – needs human sign-off
Build Files: Create directory website files (build from templates, do SEO groundwork).	95% Working
Deploy to Host: Deploy the built directory site to the hosting environment.	WIP
Connect Externals (Optional): Connect external services like Google Analytics, Google Search Console, Stripe, or set up GAlerts.	Todo
Final Output: Summarize success with a live site and any relevant access details.	Todo
Fail Output: Handles any errors or dead ends by reporting failure.	Todo

AI Challenge: Accepting Human Input

On starting this challenge I envisaged a totally human-free automated business. That might still be possible, but first we need to accept that human input is going to be part of the loop.

As I mentioned in the Blending AI post, I think the rational approach here is to build structure so that AI does nearly all of the work, but reports back and asks for sign-off.

For example my AI logo generating agent is good, but it’s not great enough that it knows which logo will work best.

For now I’ll have it generate 10 options and let me pick.

I want to make a system to record my decisions/feedback when an AI agent does ask for input, so that later we have something to work from and can automate away these steps.

Architect API: Ring-fenced Tooling

I now see the future of automated business as each company essentially being an API.

I wrote my own simple API (and named it Architect) for the first agent. Since then I’ve found that it’s a great way to:

1. Provide structured, consistent completion of some tasks
2. Ring-fence some capabilities, and so create company IP

I recommend that you too make your own API if you are following along; and I intend to share the code with my loyal subscribers if that’s useful?

API’s and Agents

As part of this blended approach I am mixing my own code-driven actions (via API), and external tool agents. I suspect this will give me the best of both worlds.

As AI agents get better at using tools (like my API), and protocols for doing so are established (like MCP servers), this modular approach seeks to balance exposure to changeable external tools against consistent ring-fenced outputs. 

Modular agents are going to be key.

I intend to expose them via this API.

Reflections on My First two AI Agents

I’m happy with the progress here, but on reflection, it’s not all gone super well:

• Workflows are messy, need standardisation
• It’s taking a lot of human setup work
• Blending AI tools is going to be essential
• I realised that making web directories is not particularly meaningful; but it is a good vehicle to test and practice all these skills
• Reaching the limitations of AI – hallucinations and the fuzzy nature of LLMs mean we currently still need logic like ‘self-healing’ and ‘retry until we get the answer we want’
• We need some form of structure or organisation platform
• Fighting the urge to make my own tooling everywhere

From that I’m refocusing.

This is an experiment to get to know the tools; the higher the percentage of automation the better, but right now it’s kink-finding not breezy automation.

What’s Next

Now that these first two agents are semi-operational, (granted they’re like two cars held together by duct tape and dreams), I want to take a bit of time to find, or build out what I need to make them run smoothly.

I’m going to explore:

1. Standardising Workflows: Expanding upon my start with FlowSpec
2. Build a formal ‘help me human’ system which records human responses
3. Try more tools, searching for a structural tool which can unify these parts
4. … and if I don’t find a good unifying platform, build one

I hope by working on standardisation, unifying systems, and formally managing human input I can stack the deck in my favour for later experiments.

It’d be amazing if I can get AI businesses to build themselves; but that’s not possible without these tools.