Most people hear “one-person company” and imagine one exhausted founder doing product, sales, marketing, support, finance, and emotional regulation from the same laptop.
That is not the goal.
The goal is to build a company where one human makes the important decisions while AI handles more of the research, production, coordination, verification, and recurring work.
Not one person doing every job. One person designing the system through which the jobs get done.
For the past few weeks, I have started building exactly that kind of company. There is one human employee.
Unfortunately, it is me.
The rest of the company lives across a stack of agents, models, markdown files, scheduled jobs, review gates, source banks, code repositories, and one increasingly consequential Obsidian vault.
Hermes is the operating layer. Obsidian is the company memory. Claude Fable 5 and Opus 4.8 handle different kinds of deep work. I sit in the middle making decisions, rejecting things, changing my mind, and occasionally staring at a terminal like it has personally betrayed me.
I am not doing this because I think AI has eliminated the need for companies, teams, or other humans. It has not. Customers remain inconveniently real. Products still break. Distribution still matters. Taste cannot be installed with `npm`. Legal, finance, support, trust, and judgment do not disappear because a model can finish a code migration at 2:14 a.m.
But the minimum viable company has changed.
One person can now hold much more operational surface area than one person could before. You can research a market, shape a product, build software, run tests, maintain institutional memory, create content, monitor competitors, synthesize customer signals, and operate recurring workflows without hiring a small department before you have earned the right to have one.
That is the part I am testing. This article is the practical version… the stack, the files, the agent jobs, the operating loops, the review gates, and a four-week plan you can use to build your own.
And I am testing it while building something in my late nights and weekends that I know all too well… a platform for becoming the answer in the evolving agentic web.
I am not ready to do the full reveal yet (more to come). The short version is that I think the web is moving from pages people browse, to answers machines assemble, to agents that research, compare, shortlist, and eventually act for us. Companies will need more than an SEO dashboard telling them they are invisible. They will need a self-learning and compounding system that helps them create the structured, proof-backed content that makes them understandable and selectable in the first place…
That is the bet.
The stranger bet is that I can build and run the first version as a one-person company. You do not need to be building a comprehensive platform to use the system. It works for a software company, consultancy, media business, agency, paid community, research product, or any small company where the founder is still the main source of direction and judgment.
A one-person company does not mean doing everything yourself
The phrase “one-person company” can summon a fairly depressing image… one sleep-deprived founder doing product, support, sales, marketing, finance, design, and emotional regulation from the same laptop.
That is not a company. That is a hostage situation with Stripe.
The AI-native version is different.
There may be one human decision-maker, but there are many execution surfaces. The founder’s job shifts away from personally producing every artifact and toward designing the system that produces, checks, remembers, and improves the work.
You still do plenty. More than plenty. The difference is where your time goes.
Instead of drafting every market brief, you decide which market question matters and what evidence would change the strategy.
Instead of hand-writing every feature, you define behavior, boundaries, acceptance criteria, and the proof required before anything counts as done.
Instead of starting every piece of content from a blank page, you maintain the claims, customer language, research, taste, and product truth the content must come from.
Instead of trying to remember what happened three Tuesdays ago, you make the company leave a trail.
This is not “AI does everything.”
It is closer to… AI handles more production while the human stays inside the decisions.
The founder becomes a strange combination of CEO, editor-in-chief, product owner, systems designer, and quality-control department.
Which, to be clear, is still a lot of jobs. But it is a different kind of lot.
The four-layer one-person company stack
The useful part is not the specific apps. You could replace pieces of this stack and keep the architecture.
The architecture has four layers:
| Layer | My tool | Job |
| Judgment | Me | Direction, taste, priorities, tradeoffs, approvals |
| Orchestration | Hermes Agent | Runs jobs, uses tools, schedules loops, routes work, keeps operating context |
| Memory | Obsidian | Holds product truth, research, decisions, proof, customer language, and output history |
| Deep work | Claude Fable 5 + Opus 4.8 | Research, synthesis, product thinking, coding, review, and long-running execution |
Each layer solves a different failure mode.
Claude without memory becomes brilliant amnesia.
Obsidian without an operator becomes a beautifully organized archive nobody uses.
Hermes without clear standards can automate the wrong thing at impressive speed.
And the human without any of this becomes the bottleneck, the database, the project manager, the reviewer, and the person wondering why “working for yourself” feels suspiciously like being employed by every department at once.
The stack works when the parts are separated on purpose.
Obsidian is the company brain
Most people begin their AI workflow with the model.
I think the more important place to begin is the memory.
My Obsidian vault holds the company’s durable context: the product thesis, positioning, architecture decisions, source research, competitor notes, customer language, build logs, claims, proof, open questions, editorial plans, and the things the system is explicitly not allowed to say or do.
It also separates different kinds of truth.
- Raw sources stay raw.
- Assumptions are labeled as assumptions.
- Product decisions have reasons and dates.
- Public claims need proof.
- Private context stays private.
- Shipped work feeds the next round of work.
This sounds boring because… well it is boring.
It is also the sole difference between an agent helping run a company and a chatbot giving you a confident answer based on whichever paragraph happened to be nearest.
A company has memory whether you design it or not. Usually that memory is spread across Slack, somebody’s head, old decks, half-finished Notion pages, email archaeology, and one person named Dan who apparently knows why the pricing page says what it says.
A one-person company cannot afford that kind of institutional fog. There is no Dan. I checked.
So the context has to live in the system.
For me, the vault knows what the product is trying to become, which language is canonical, which product decisions are binding, what has already been built, what still needs founder judgment, what evidence supports the agentic-web thesis, and what should happen next.
That means a new AI run does not need me to reconstruct the company from memory every time. It can inspect the relevant notes, distinguish source material from synthesis, and enter the work with context.
The point of the second brain is no longer only helping me remember.
It is helping the company remember.
Hermes is the operating layer
Obsidian stores the company’s knowledge. Hermes turns that knowledge into motion.
Hermes can inspect files, operate tools, run code, use the browser, schedule recurring jobs, delegate contained work, update the vault, and bring the result back for review. More importantly, it can preserve the operating rules around the work.
That changes the unit of interaction.
I do not have to ask:
Can you summarize these five competitor pages?
I can define a recurring competitor-intelligence job that:
1. Checks an approved set of sources.
2. Captures what materially changed.
3. Separates facts from interpretation.
4. Explains why the change matters to Citaeo.
5. Saves the dated signal in Obsidian.
6. Updates durable strategy only when the evidence warrants it.
7. Stays quiet when nothing meaningful happened.
That last part matters more than it sounds. A useful AI company does not need more bots announcing that they completed their little chores. It needs systems that know when there is something worth interrupting the founder about.
I use Hermes as the connective tissue between the company brain and the work:
- research gets routed into source notes;
- product decisions update the right context;
- recurring scans create dated evidence;
- weekly reviews distill what changed;
- content drafts pull from current product truth;
- completed work leaves behind decisions, proof, and next actions;
- risky or public actions stop for human approval.
The goal is not maximum autonomy.
The goal is useful autonomy with receipts.
Claude does the work that needs depth
Not every job needs the same model or the same amount of thinking.
For quick operational work, the smartest possible model is often unnecessary. You do not need to summon a digital philosopher-king to rename a file and check whether a link works.
But some jobs are large, ambiguous, and deeply connected. They require reading across a codebase or a set of strategic documents, making a plan, holding constraints in mind, challenging assumptions, executing over time, and proving the result.
That is where I use Claude Fable 5 and Opus 4.8.
Fable 5 is useful for the sprawling jobs: inspect the system, understand the goal, work across many files, create or update artifacts, verify the output, and preserve what the next run needs.
Opus 4.8 is where I want more judgment density: architecture, difficult product decisions, strategy, synthesis, writing review, and the places where a merely plausible answer is expensive.
I do not treat either model like an oracle. I treat them like high-capacity workers that need a strong job brief.
A real job brief includes:
- the outcome;
- the source map;
- the constraints;
- the decisions reserved for me;
- the verification standard;
- the files or systems that should be updated;
- the conditions that mean the work should stop.
“Build this feature” is a request.
“Inspect these product docs and code paths, restate the expected behavior, identify conflicts, propose the smallest useful implementation, write tests first, build it, run the test suite and type checker, exercise the actual flow, and return the changed paths plus remaining risks” is a job.
The model did not become more intelligent because I used more words.
The company became more legible about what done means.