Harlem Labs Advisory · Field Note

The infrastructure is building itself.

The era of gluing together single-purpose tools is in its twilight. The new operational reality is agents that build the systems that do the work — and the key artifact in that reality is a plain text file.

April 2026 · 7 min read
The confession

Years of work disappeared in days.

As I look at my screen, I struggle to process the compression of time. If you had asked me a month ago how long it would take to execute the backlog of operational debt sitting on my desk, I would have measured it in quarters. Not weeks. Quarters.

I run a Strategic AI advisory firm, which means I spend my days advising others on how to structure their systems. Behind the scenes, like any operator, I had my own mess to clean up — years of accumulated folders, disconnected tools, and infrastructure that had grown by accretion rather than design.

Over two weeks I reorganized my entire digital workspace. I collapsed scattered materials into structured documents, connected them to Notion, and started clearing out Airtable. With that infrastructure finally clean, I turned to the Harlem Labs website and split it into two distinct properties, both being built with entirely new tools.

And then, because the operational weight was gone, I had the free time to sit down with my son and help him build a video game.

That is the confession. I am looking back at years of work that disappeared in days, and the reality of what happened is still settling in. But the real story is not about getting organized. It is about how it happened, and what it reveals about the way we run businesses now.

The illusion

The stack was never the product. The overhead was.

For years my stack looked busy: n8n, Manus, ChatGPT, Airtable, Bluehost, WordPress, assorted plugins, GoHighLevel, Google Forms, and a host of others. I had accepted the friction of connecting these systems as the cost of doing business. I spent my time maintaining the machinery instead of driving the vehicle. Every tool solved one problem and created three more.

I built entire workflows and job descriptions around managing that friction. I hired specialists to configure my CRMs, developers to maintain my WordPress sites, and an assistant to make sure the administrative work got done and communicated between platforms. I told myself this complexity was necessary — that it was the price of running a real business.

It was not. It was the limit of the technology at the time.

GoHighLevel is gone entirely; what I built in it now lives in Airtable and in tools built directly by AI. Bluehost is being replaced by Vercel. WordPress is gone for good. Even my forms are built in a terminal now. Not because those are bad products — because I no longer need them. The migration did not take weeks of planning and careful handoffs. It happened as a byproduct of the market shipping tools that build the tools. That distinction matters.

From prompts to agents to skills

Three phases, arrived at in a hurry.

Phase one
The prompt
Ask a model to write the email, summarize the document, draft the code. A parlor trick that became a utility — but a human still had to execute the output.
Phase two
The agent
Give the model tools. Let it search, run code, execute multi-step work. We stopped asking for the recipe and asked it to cook. This is where most operators are right now.
Phase three
The skill
A plain text file that encodes a process, a workflow, a set of rules. Codified operational intelligence. The tool becomes the description of the tool.

The deeper insight took me a few days to articulate: AI has not only been my executor. It has been the thing that builds the things that build the things that do the work. It did not just do the filing. It built what WordPress was solving. It built what GoHighLevel was solving. The infrastructure layer itself is now generated.

The codified business

A map of operations.

When you replace a rigid tool with a described process, you change the economics of execution. This is my own stack, before and after, over those two weeks. If you run a small or mid-sized business, you should see your own operations somewhere in it.

FunctionWhat it used to costWhat runs it now
Website and CMSWordPress, Bluehost, a stack of slider pluginsA site built and deployed straight to Vercel. Faster, cheaper, fully under my control.
CRM workflowGoHighLevelStructured logic running inside Airtable, built and maintained by an agent.
Forms and routingGoogle Forms and a paid form serviceCustom forms built in a terminal, deployed with the site itself.
Lead qualificationVendor logic I rented and could not readA described rule set an agent runs — and that I can open and edit.
Onboarding sequenceA sequence tool with a monthly seat priceA described process in a file. Version-controlled, portable, mine.
Content pipelineThree tools taped together, plus the person taping themOne file that says what finished looks like, run on demand.

Every row here is a process I could already describe out loud. That is the whole qualification. Nothing on the right side required me to write code.

The evidence

Forty percent is not an incremental gain.

In a 2023 field experiment with 758 consultants at Boston Consulting Group, researchers from Harvard Business School and MIT Sloan found that when AI is used within the boundary of its capabilities, it improves a worker's performance by nearly 40% compared with workers who do not use it.

38%
performance increase for the AI-assisted group over the control
Harvard Business School / MIT Sloan, 758 BCG consultants, 2023

That is a restructuring of what one person can accomplish in a day. When the infrastructure builds itself, that gain is no longer spent managing the stack. It is spent executing the strategy. The advantage that once required headcount can now be built with clarity of process and the right tools to run it.

Test your own operations

Three questions.

1
Which tool are you paying for monthly to do something you could describe in a paragraph?
2
Of the processes you run every week, how many are written down anywhere other than in your own head?
3
If you handed your operation to someone tomorrow, what would you have to sit beside them and explain? That is the list.
The other side of the threshold

What reclaimed capacity actually looks like.

The two sites are not finished. Airtable still has most of its cleanup left. There is more to explore in these systems than I have had time to test. But the threshold has been crossed, and I am still catching my breath.

I spent years building systems for other people, always aware of the operational drag that came with them. That drag is gone. The systems are building the systems.

And the most significant result of all that compression was not a better spreadsheet or a faster website. It was the afternoon I spent with my son, watching him build a video game, knowing the machinery of my business was quietly running itself in the background. Not a dashboard metric. Not a productivity report. An afternoon that would not have existed six months ago.

You do not buy the tool anymore.
You write down the process clearly enough to hand it over.
Map → Build → Evolve. Capability, not dependency.

Describe it once. Run it indefinitely.

Most operators do not need another platform. They need their own processes written down clearly enough that something else can run them. That is the work we do.

Reference

Dell'Acqua et al., “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality,” Harvard Business School Working Paper 24-013 (2023). hbs.edu

Harlem Labs Advisory
Strategic AI infrastructure for operators