For the last couple of years, we’ve been building custom GPTs. You train an agent to do SEO, or audit Facebook ads, or manage analytics. You connect it to your CRM, WordPress, Google Ads, and the rest. Then you share that agent with your team through a business OpenAI account.
That worked. But it meant everyone depended on a centralized agent, and scaling it was clunky. I’ve come up with something more elegant.
The three-layer system: Task Library, definitive articles, and meta articles
Everything starts with the Task Library, which contains over a thousand tasks. Each task has a definitive article, which is the master SOP defining step by step how to get that task done. Attached to each definitive article is a skill.md file that agents can reference when executing the task.
When an agent executes a task following the definitive article, it generates a meta article documenting exactly what it did for that specific client or instance. Meta article 1, meta article 2, meta article 3, and so on. The more times a task is executed, the more meta articles accumulate under that definitive article.
This is the part most people get wrong: the meta article is not the same as the definitive article. The definitive article is the recipe. The meta articles are the documented instances of a cook following that recipe, noting what happened each time.
The recursive loop (recursive self-improvement)
The recursive loop is not a single component. It’s an iterative cycle across multiple steps:
- An agent executes a task following the definitive article.
- The agent generates a meta article documenting what it did, how many steps it took, how many tokens it used, what inputs were needed, and what decisions it made along the way.
- As meta articles accumulate, patterns emerge: where agents trip, what inputs produce better results, what steps are unnecessary.
- Those insights feed back into the definitive article, updating and improving the master SOP.
- The next time any agent executes that task, it follows the improved definitive article.
This is recursive self-improvement (RSI). The system gets smarter with each execution without a human manually updating a thousand SOPs. Anthropic themselves reported generating eight times more code than a few quarters ago because of this same principle.
Why this works for SEO
Most local businesses have no link juice. They can create content all day, but without signals flowing to it (links, citations, entity connections), that content is practically invisible.
Our agency site, BlitzMetrics, carries a DR63. When we document work we’re doing for clients through meta articles and cross-link between sites in our network, we’re passing legitimate authority to sites that have none. This isn’t black hat PBN link buying. It’s real work, documented transparently, following Google’s search quality rater guidelines and EEAT standards.
Take one of our roofing clients with multiple locations under a single mothership brand. The mothership has power, and we drive that to each location through intelligent cross-linking. The same logic applies to personal brand sites we build for people like George Leith, who scaled Vendasta to a billion-dollar company, or Brady Sticker, who runs Church Candy, an almost eight-figure agency serving churches.
Every meta article strengthens the knowledge graph by clarifying what the entities are, the objects associated with those entities, and how they connect to each other.
The progression: prototype, tune, then scale
A critical mistake I see people make is jumping straight to persistent agents without prototyping first.
The correct progression: start in the browser (Chrome extension, Atlas agent, or similar) where you’re already logged in and can demonstrate tasks to the assistant. Do the task yourself at least three times, tuning the skill files each time. Only after you’ve validated the process do you move to persistent agents (like CoWork) that can run on a schedule without you watching.
Skipping prototyping and scaling something that isn’t ready creates a mess. You’re just multiplying errors.
How I actually use this
I had a Zoom call with Erik Huberman of Hawke Media, a $700 million a year agency. Erik has acquired 23 agencies and wanted to talk about how he’s winning in spite of AI.
I told my agent: go to my Zoom recordings, grab the transcript, pull screenshots from my Google Photos of Erik and me over the years (Miami, Austin, San Diego, Santa Monica), and turn the whole thing into an article on blitzmetrics.com following our article guidelines.
I gave it context about Erik’s track record, how he keeps the people at companies he acquires, how he’s a growth partner rather than a PE firm gutting companies. That task would take a human hours. It took me a few minutes to give instructions. Then I moved on to the next thing while the agent worked for an hour or two across Zoom, email, Google Photos, and WordPress.
Two things made this work: I referenced our definitive article on how to repurpose a Zoom into an article(so it followed our SOPs, not random internet advice), and I provided entity context about how Erik and I are connected (so the resulting content has real EEAT signal).
The competitive moat
Here’s what most people miss about RSI: the accumulated execution history IS your competitive advantage.
Without these loops, anyone can ask their AI to reverse-engineer what you did and beat you. The moat isn’t the SOP itself. It’s the accumulated execution history.
You can open-source your skill files. Facebook open-sourced the code to run their social network. They weren’t worried because the real value is in the data, the relationships, and the proof of execution.
When Jennifer (our article grading agent) has done something thousands of times and has thousands of meta articles documenting that track record, that experience can’t be replicated by someone who just grabbed the skill.md file. The more times an agent executes a task, the stronger it gets. This is a corollary to LDT/CCS.
The system is also model-agnostic by design. We’ve built our metadata so we can swap the engine for Gemini, OpenAI, or Grok as models leapfrog each other every few weeks. The value isn’t in which model you use. It’s in the recursive loops you’ve built on top of it.
The end state is an agent marketplace. Each agent is priced per task based on token cost versus value delivered. If it costs a penny in tokens and is worth $5 of human effort, you charge a dollar. Costco model: fair markup, more volume, more value accruing to the buyer. This creates a repeating advantage where you make money by serving a lot of people rather than trying to extract maximum price one at a time.
The MAA framework
Years ago, when we had Quiznos as a client with 5,700 locations, I couldn’t personally talk to every franchise owner. So I created a framework called MAA: Metrics, Analysis, Action.
Instead of dumping 20 pages of screenshots from Google Ads and dashboards, MAA requires translating data into business terms. Is a 1.7% CTR good or bad? Are they profitable? What does ranking on 498 keywords actually mean for their business?
Now we have agents send clients a simple text or email every Friday following MAA. Two to three sentences: here’s what happened this week, here’s what it means, here’s what we recommend. That does 90% of what an account manager would do, and it turns every client interaction into celebrating wins rather than fielding complaints.
Why documentation is the whole game
The real insight isn’t that agents can do work. The next level is: did it document how it did the work? How many steps, how many tokens, how long it took, what input it needed, what it could improve.
Then you ask: does our definitive article reflect what actually happened? If agents keep making the same mistakes across multiple meta articles, fix the root cause by updating the definitive article. Solve it at the root, not at the instance.
This recursive system, where agents do the work, document the work, improve the process, and make that improvement available to all other agents, is how you actually scale. The fire either refines the gold or it burns off the chaff.