What Comes After Agents

For the past few years I built agents — browser agents, deployment agents, function-calling primitives, copilots. I got good at it. And I kept walking into the same wall, which I will state plainly because it took me too long to admit out loud: a more capable agent did not give me a more capable system. Each new model made the demo better and left the real work almost exactly where it was.

The work that never budged was never the part inside the model. It was everything around it. An agent that can write the code cannot tell you whether the code should be written, whether someone already tried it last month, who will be annoyed if it ships, or what "done" even means here. It has the skills of a senior engineer and the situational awareness of someone who walked into the building thirty seconds ago — every single session.

So I stopped trying to make the agent smarter. I started building the thing it was missing.

That thing is Singular — an applied research lab building AGI products for human–AI collaboration. We are building the operating system for AI-native companies: turning the scattered work and knowledge of an organization into a living company brain that AI coworkers can read, reason over, and write back to. (If you want the full argument for why the company — and not the agent — is the right unit to build, I made it here. This post is about what I'm actually building, and what I'm still unsure of.)

A company is an answer to questions an individual can't answer alone

It helps to ask what a company even is, stripped of the furniture. Most of what an organization does is not the visible work — the code, the deals, the designs. It is the invisible machinery that lets many people act as if they shared one mind: a sense of what is true right now, what matters most this week, who gets to decide, and what already happened so nobody relitigates it. New hires spend their first months absorbing this, and we call them productive only once they have.

Agents have none of it. We hand them a prompt and a blank slate and act surprised when they behave like a brilliant stranger. The fix is not a longer prompt. It is to give them what a company gives a person: a place to belong, a memory that persists, colleagues to coordinate with, and a clear line for when to ask a human. Build those well and the agent's raw capability — which was never the problem — finally has somewhere to land.

A company with a brain

The foundation is memory, and I want to be precise about what I mean, because "give the AI memory" has become a hand-wave.

I do not mean a vector database bolted on at the end so the model can recall a paragraph. I mean the substrate the whole system runs on: the documents, decisions, conversations, and tasks of a company, turned into something structured and queryable that every coworker reads from and writes back to. The point is that context stops resetting. It compounds. A company that has operated for six months should beat one that started yesterday for the same reason a six-year employee beats a new hire — it remembers what was tried, what failed, and why.

This is also the hardest part, and I am not going to pretend otherwise. Human institutions spend enormous effort keeping their shared memory from rotting — deciding what to write down, what to forget, what to revise when the world changes, how to resolve two records that disagree. A company brain that merely accumulates becomes a hoard, not a memory. Getting that curation right — what compounds versus what is noise — is most of the actual research.

The product takes shape around four pieces, and I think of each as the answer to one of those questions:

• AI Coworkers — Who does the work? Not stateless functions you call, but employees with persistent memory, skills that grow with experience, and their own expertise. You recruit them, and they get better at the job.
• Company Emulator — How do they coordinate? You design departments and teams that self-orchestrate as a swarm over long horizons, instead of a human routing every handoff by hand.
• The Workspace — How do you stay in control? Meeting rooms, whiteboards, and calendars where you can actually watch the company work and step in. Oversight as a seat at the table, not a setting in a config file.
• Toward Singularity — Where does it go? A path from human-guided work toward genuinely autonomous operation, powered by coworkers that improve themselves and each other.

Where the human stands

I do not believe the interesting version of this is "fire everyone and watch the robots." The interesting version is a slow, deliberate change in where the human stands. You start in the loop — approving, correcting, teaching. As the company brain fills in and the coworkers earn trust, you move onto the loop — setting direction, watching the workspace, intervening only when it matters.

The trap is that "watching" quietly becomes a fiction: a wall of green checkmarks over a system you no longer understand. Keeping supervision real as autonomy rises is, to me, the deepest design problem in the whole thing — and the reason Singular is a research lab and not just a product. The endpoint we are studying is the boundary itself: how far can a company run on its own, and what does a human's job become when it does?

What I'm not sure of

Conviction is cheap, so here is the honest ledger. I don't know whether "fully autonomous" is ever actually the right setting, or whether the best companies will keep a human firmly on the loop forever — and I suspect the answer differs by domain. I don't know whether a swarm of specialized coworkers truly beats one very strong generalist agent for a given task, or only feels better because it mirrors how we organize. And the memory problem above is genuinely unsolved: compounding context could just as easily compound mistakes. I am building Singular partly because I want to find out, not because I already know.

Why now, and why this

Three things finally lined up: models good enough to act, memory and long-context techniques mature enough to give them a history, and orchestration patterns good enough to let many of them work together without dissolving into chaos. None of these alone is enough; together they make an AI-native company buildable for the first time. The agent was the prerequisite. The company is what comes after.

If you care about human–AI collaboration, autonomous organizations, or you just want to recruit your first AI coworker, come find me — or look at singularos.com. The company brain is online. Let's see how far it can run.