We Experimented a Prototype Business OS for a Month. Our SEO Grew 10x.

In March of this year, I started an experiment: I wanted to find out what it would take to build the prototype of a Business Operating System for a small 1–3 people startup like Epsilla. Not a strategy doc, not a slide deck about what AI agents could theoretically do. A working system that actually runs the business.

I picked one workflow as the first test. SEO content for our official website. We had been publishing two or three articles a month, sometimes less. I wanted to see if we could push that to five or ten articles a day, without hiring anyone, without writing a line of code, and without losing voice.

One month later, here is what Google Search Console looks like.

Google Search Console, Feb 7 to April 4, 2026. Clicks up roughly 10x. Impressions up roughly 36x. Average position improved from low teens to around 7.

I want to explain how this worked, what it is still bad at, and why I think this is the early shape of something much bigger.

Why 2026 is the year of the Harness

A bit of context first.

For most of the last two years, AI agents have been a conversation about models + ReAct loop with function calls / tool uses. Which model is smarter. Which model can call tools. Which model can write better code. The agent itself was a thin loop wrapper around a chat completion.

That started changing since Claude Code launched last year, and became a mainstream framing when OpenClaw getting serious adoption earlier this year. OpenClaw is not a simple ReAct loop agent. It is a harness. It is the runtime that turns a model into an actual coworker, one that runs on your devices, sits in your messaging channels, watches for signals, and accumulates context over time. The same shift happened in parallel for Hermes, for Claude Code, and for the new generation of Codex.

I think 2026 will be remembered as the year the harness became the product and commoditized. The model is the engine. The harness is the car.

Harnesses change what an AI agent can do for a small company in three specific ways (the original framing is from https://www.superlinear.academy/c/aa/).

Frictionless. OpenClaw and the others put the agent inside the messaging channels you already use. Telegram, WhatsApp, Slack, iMessage. You don’t open a separate app to work with the agent. You text it. You voice-message it. You CC it on a thread. The friction of context switching drops to almost nothing, and for the first time it actually feels like collaborating with another person rather than operating a tool.

Proactive. OpenClaw has a heartbeat. Claude Code has a /loop. The agent is not waiting for you to ask. It is running on its own clock, scanning for the things you told it to scan for, and pinging you when something matters. This is a quiet but enormous shift. Software used to be reactive. Now it can be the participant that starts the conversation.

Contextual. The agent shares a single, persistent context across every session. Anything you have ever told it, anything it has ever observed, anything the company has written down becomes part of its working memory. Work stops being fragmented across throwaway sessions. Every conversation builds on every previous one. The institutional knowledge of the company stops living in scattered Notion pages and starts living in something an agent can actually read and reason over.

These three shifts together are the foundation of what I think is the most important new product category of the decade. The Business Operating System.

The Company Brain

Y Combinator have been talking about a concept they call the Company Brain. The shape of it is the same idea from a different angle. Take everything a company knows, all of its internal communication, all of its external signals, all of its product context, and unify it into one place an AI can read from and write to. Then point a harness at that brain and let it run the operating loop.

[embed]

GBrain by Garry Tan: https://github.com/garrytan/gbrain

That is the vision. The question is how you actually build it.

In March I decided to find out, with the smallest possible experiment.

The three layers of a Business OS

The way I think about the Business OS, it has three layers.

The context layer. Everything the company knows, in one place. Documents, code, customer interactions, strategic notes, past decisions, current projects, a dump from ChatGPT conversations, Notion, Google Workspace, Slack, Linear, Github, everything. The brain.

The harness execution layer. The runtime that lets agents act on that context. OpenClaw plays this role for us. Hermes, Claude Code, or Codex can be used as the same thing. It is what executes the work and talks back through the messaging channels we already use.

The business workflow layer. The actual jobs to be done. Marketing, sales, customer support, product development, finance. Each one a workflow that sits on top of the harness and the context.

You can build any one of these in isolation, and people do. The point is that all three together is what makes it an operating system rather than a tool. The context layer without a harness is a knowledge base. The harness without context is a chatbot. Both without a workflow layer are demos. The combination is something new.

The experiment

We picked SEO content as the first workflow. Concrete enough to measure, contained enough to ship, important enough to matter.

Context layer.

I created a folder called ClawWork on my machine. Into it I put everything. Our project repos, our product specs, our internal strategy notes, the YC bookface materials I had been collecting, transcripts of customer calls, market analysis I had written for myself. All of it sat in one place, organized as files. OpenClaw was given access to that folder as its primary context.

I also gave the brain a name and a personality. I called her ElizaClaw, my AI Co-Founder with INTJ personality. The reason for naming her matters more than it sounds. A personality forces consistency in how you write to her and how she writes back, and that consistency is what lets you build a working relationship with her over months. Every day I check in with ElizaClaw. Every day she scrapes Hacker News, Google News, and a handful of niche AI sources I have configured, summarizes what she found, and writes it back into the context folder. The context is not static. It is compounding.

Harness execution.

OpenClaw runs in the background on my machine. It is connected to Telegram, so I can talk to it from anywhere using my phone’s voice input. It has the keys it needs to do work. GitHub access. Vercel deploy credentials. The SEO plugin stack. Image generation API keys. I never wrote a deploy script. I described what I wanted, and OpenClaw figured out the rest through several rounds of trial and error.

Workflow layer.

The actual SEO workflow has two flavors.

The first is automatic. Every morning, ElizaClaw pulls the day’s signals from her sources, identifies the developments that intersect with our product positioning, and drafts a publishable blog post about each one. Most days that is three to five articles. They go through the SEO plugin stack, get illustrated, get a slug, get pushed to the GitHub repo, and get deployed via Vercel. By the time I am up, they are live.

The second flavor is on demand. I see something on X or read a paper, I think it deserves a longer piece, and I voice-message Telegram with my angle. ElizaClaw drafts the article in our voice, runs the same plugin stack, and ships it. Total time from idea to live post is sometimes under 10 minutes.

Both flavors share the same context, the same harness, and the same deployment pipeline. The only difference is whether the trigger came from a daily signal or from me.

The results

Over the first month, we went from publishing two or three articles a month to publishing five to ten a day.

Google Search Console tells the rest of the story. From February 7 to April 4, total clicks grew roughly 10x and impressions grew roughly 36x. Average position climbed from the low teens to around 7. Signups followed.

The blog screenshot at the top of this post is page seven of thirty seven. Each page has six articles. That is the inventory we built in 30 days, with one human in the loop and no code written by hand.

What this is not

I want to be honest about what we have built. It is a prototype, and it is rough.

If I am being blunt about the output, most of what we publish right now is AI slop. Synthesized news. Competent surface-level analysis. Decent SEO content. Far from the deep original insight I want this brand to be known for.

I want to be precise about whose fault that is. It is not a limit of the technology. The harnesses can do better. The models can do better. The signal is out there in the world. What is limiting the quality is us. We are still learning how to direct the system. We are still developing the skill of writing prompts that elicit genuine insight rather than competent summary. We are still figuring out what the human’s role looks like when the drafting is delegated and only the directing remains. The bottleneck is our taste and our skill, not the stack. We are iterating on this every week, and the quality bar a month from now will be higher than the quality bar today.

The Telegram interface is also primitive. It works because I have stopped resisting voice input and because OpenClaw is patient with my edits, but it is not the polished surface that a real productized version would have. Half the time I am sending stream-of-consciousness messages and trusting ElizaClaw to make sense of them.

A productized version of this system would also need to solve things we currently solve by hand. Better feedback loops from analytics back into the writing prompts. Better signal scoring so the agent ignores noise. Real editorial review by another agent before publish. A polished interface that does not depend on a single founder being patient with voice transcripts. All of that is the work ahead.

What’s next

The blog workflow is just the first one. The point of an operating system is that once the context layer and the harness layer are in place, every additional workflow gets cheaper to build. The marginal cost of the second workflow is much lower than the first, and the marginal cost of the tenth is almost nothing.

We are already wiring up the next ones. Customer interactions and support. Marketing campaigns and email sequences. The early loops of our sales process. Product development, where AI coding agents become another participant in the harness. The pattern is the same in each case. Identify the workflow. Wire the harness to the context. Let the loop run. Watch what breaks. Fix it. Compound.

The eventual product is not a single tool. It is a Business Operating System sized for very small teams and one-person companies. Our handcrafted prototype is the early version of something we plan to give away as a product to tens of millions of new one-person companies that will be founded by the end of 2027.

For now, here is the only insight I am confident in. The shape of an AI-native company is not a small company with AI tools bolted on. It is a small company with a brain. Anyone reading this in 2026 who is not already building toward that shape is going to look up in twelve months and discover that the operators who started earlier have a multiplier they cannot match.

If you are trying to figure out how to start, my advice is the one I followed. Pick one workflow. Wire the simplest version of the three layers around it. Ship the prototype. Iterate from there. The first month is the most important month, because you will learn more about your own business by watching an agent try to run it than you will from any consulting engagement.