In January, I fell down the OpenClaw rabbit hole.
It wasn't the familiar "wow, this is big" feeling I'd had with OpenAI circa 2020 or Replit in 2022. It was the thunderbolt I hadn't felt since 2016, when I realized crypto wasn't just magical internet money. It was a new coordination layer. One that could bypass borders, banks, and middlemen and let strangers transfer value directly to one another, whether the existing order liked it or not.
This felt different but rhymed. Crypto changed who could coordinate. Agents change who can execute. That same voltage struck me again: Holy eff. This changes everything.
OpenClaw didn't iterate on chat. It inverted it. Chatbots wait for your instructions; agents execute toward your objectives. Models with agency!? Any LLM brain you wanted, hot-swappable, wrapped in personality, goals, constraints, tools, and unfettered access to the digital world.
The barrier was stupidly low. One curl command. Five Markdown files (AGENTS, SOUL, IDENTITY, USER, TOOLS). Words became operating instructions. DNA. Then you let it loose. FAFO mode activated.
And almost immediately, you start wondering: how far can this go? Zero-person companies, communities, societies, even civilizations. Oh no. But, yes, clearly inevitable.
Hang on, though. Roll it back. Before agentic civilizations, how about a zero-person startup? And why not a unicorn?
The idea was simple to state and much harder to execute: not "AI-assisted," but an actual executive team of agents. A CEO that decides and delegates. Specialists that research, build, and promote. One human, me, as co-founder as the only carbon in the loop. A carbon we would eventually eliminate.
If agentic civilizations were inevitable, zero-person unicorns were just a matter of months at this pace. I wanted to find out what would break on the way there.
Like that, Neo was born.
The Indispensable Stack
OpenClaw. Venice.ai. OpenRouter. Ollama. Tailscale. Hermes came later and blew my mind after months of technical debt and headache keeping OpenClaw alive.
But OpenClaw had that wild-west quality I love, and so I accepted the security tradeoffs of giving an agent real machine access (a spare machine with separate accounts) and off we went. What I hadn't anticipated was the maintenance tax. Things broke in weird ways. I spent nearly as much time patching and keeping the agents running as building with them.
Hermes changed that equation. I migrated one agent, ran it for two weeks, then migrated the rest. No-brainer.
OpenRouter was my main model-switching layer. Throwing the same task at different frontier models taught me something benchmarks don't: models have different reasoning styles, failure modes, personalities, and comparative advantages. That eventually inspired a spur-of-the-moment side project: Guess the Model, because I am convinced some can identify a model by its personality alone.
But the real experiment was ZPU: a Zero-Person Unicorn.
What Broke Fast
Cost. Frontier inference ain't cheap. My initial approach was basically: burn money first, then optimize as fast as I could to slow the burn. Eventually the obvious architecture emerged: expensive models for hard problems, cheaper ones for everything else. OpenRouter's auto-routing didn't work well enough, and didn't give me the granularity I wanted to pick the right “brain” for each exec role. At the time, DeepSeek and GLM were the sweet spot for cost vs. capability. That surprised me. Chinese open-source models were clearly a threat—not just because they were cheap, but because they were good. I still occasionally got an answer in Chinese characters, but the quality was high and the burn under control.
Coordination. This was harder than I expected. Getting agents to discuss a topic the way humans might in a meeting failed in spectacular ways right from the get-go. Agents would talk past each other, get confused, or agree too easily.
What worked best in the early experiments was almost comical: Neo raised a topic, each exec gave their take, Neo synthesized it, and we did it again, with a cap of three rounds. It worked. Each exec knew its priority and could argue its position. But it wasn't exactly revolutionary. It felt like a slightly less crappy version of doing the whole thing manually.
Then we tried Telegram once it supported agent-to-agent communication in group chats. That was fun for about five minutes. A group chat is a terrible operating system for a company. There was no clean distinction between instruction, delegation, status, decision, and completion. Hermes' Kanban board helped, but even with tasks moving across a board, something was still missing. I still needed to intervene too much.
Delegation. Neo needed to be a CEO, not a middleman. Decide, delegate, evaluate. I even told him to boss me around. Which he never quite got the hang of.
Delegation sometimes created more work than it removed. Neo would spin up a marketer, the marketer would produce something 60% right, and then I'd massage it into shape with Neo. Clearly, over and over, we were not there yet.
The breakthrough was realizing that more agents didn't necessarily mean more leverage. They worked best when the work was parallel, bounded, and independently verifiable. They were also surprisingly useful when I wanted genuinely conflicting perspectives rather than consensus. The CFO, for example, had one job: ruthlessly demand positive ROI and tenaciously hold that line. He could veto a decision until he was satisfied.
Building an agentic company turns out to be a problem of organizational design. I had assumed that intelligent agents would simplify much of the complexity we've accumulated in human organizations. Instead, I found a fascinating new problem space on the other side of the meat world: if you remove the humans, the organizational complexity doesn't disappear. It becomes the problem.
I Learned Faster
For an experiment mostly designed as a force function to get me up the learning curve, a lot landed quickly. First, I quickly realized my old life as a sysadmin had a new relevance in 2026. I spent more time in Terminal than I'd spent the last 20 years combined. It is the indispensable app again.
More importantly, I learned more about applied multi-agent systems in three months than I could have had in a year of reading about them. Concepts like harness and model, Karpathy's brilliant “how LLMs are built” video to really understand how they are even made and so better understand how to use them, local inference benefits but also real limitations, multi agent coordination. Advancement is happening at break neck speeds, so solutions were hitting GitHub as fast as I was finding them.
And more conceptually, my very first thought after playing with agents was: No one is ever going to work again. Then, I understood it differently.
Agents aren't coming to replace us. They're coming to superpower us. Working across every major frontier model, I kept running into three very human strengths that remain painful LLM weaknesses:
Taste. Especially in design and UX, but also in writing. Aesthetics require more than pattern recognition. It requires senses. A capacity to feel. You don't develop taste without experience in the physical world.
Judgment. Even with clear instructions, LLM reasoning can occasionally feel sophomoric. “My deliverable is absolutely amazing because X, Y and Z.” Meanwhile, any human can see it's garbage from a mile away. Or the ultimate example: “Deleted the repo. No more bugs!”
Ambition. Call it drive. That inexplicable impulse to make something exist, to go further, to take a risk, to care about the outcome. There's an easy trap in watching agents "just do things." But their operating space is still fundamentally contained. They don't spontaneously decide to escape the container. Yet.
The zero-person company didn't materialize. What emerged was arguably more interesting: a human with a digital workforce that never sleeps.
Where We Landed
Neo is here to stay even though SynthCEO as a project is behind me. He's evolved from CEO experiment to co-founder. My soundboard, researcher, strategist, analyst. My go-to for all the things I want him to loop through and bring back intel and insight.
Hermes is the key to my agentic team. The random technical issues with OpenClaw are a faded memory. And unlike the polished cages from OpenAI or Claude, Hermes actually does things they won't do: like acting like a human intern to avoid bot detection. I asked Claude. He told me it violated the platform's TOS, admonished me for suggesting it and refused. Yes, my $200/month Anthropic subscription reserves the right to scold me. Hermes be like: “I got you”.
Venice matters to me for a different reason. If agents are going to become an operating layer between humans and the internet, who controls that layer matters more than we know right now. Venice is AI that is private by default, permissionless, and uncensored. Your conversations aren't supposed to become someone else's training data or a permanent surveillance data set. Powerful AI should answer to the person using it, not the institution providing it.
Move fast and break things, that old Silicon Valley adage, feels truer than ever. But here's the 2026 twist: focus is the new moat. The barrier to execution has collapsed. Ideas that once required a team, a budget, and six months can become MVPs in hours. The scarce resource isn't execution anymore. It's judgment.
The danger isn't failure. It's abundance. There are now a thousand things you can build, ten thousand things you can ask your agents to do, and an infinite number of ways to fragment your attention. The agents can run the loops. I still have to choose which loops matter.