One Build, No Employees: An Intelligence Graph, Start to Finish, on Agents

The thing I can say that most people in venture can't: the company that just built this has no employees. NUVC runs on 24 agentic teams and zero headcount, and I want to describe one ordinary build, because the abstract version ("AI agents run the firm") is easy to nod at and easy to disbelieve.

The task was a venture-outcome graph — thousands of accelerator companies, their cohorts, founders, investors, and exits, in a production database. Here is what actually happened, in order, without a human doing the middle.

What the agents actually did

The agents designed the schema — a proper star model, dimensions and facts, with the outcome events kept as an append-only time series because that's the layer that becomes training labels later. They sourced the public data. They transcribed pitch videos when a transcript was the cheapest path to a signal, and fell back to a paid model only when it wasn't. They ran entity resolution against the existing directory. They caught their own bad matches — the two-letter-token false link, the accelerator's own cheque masquerading as a raise, the namesake company three time zones away — and corrected them. Before writing to the live database, they ran the migration as a dry-run inside a transaction, found a bug in how a UUID was being quoted, rolled the whole thing back, fixed it, and only then committed.

That last sentence is the whole point in one breath: not "agents did a task," but agents ran a process with a safety discipline — plan, dry-run, verify, roll back on doubt, commit only when it holds. The orchestration wasn't the achievement. The judgment about when not to write was.

The part people skip

I'll resist the temptation to make this sound frictionless. It wasn't. The failure modes of an agentic build are different from a human one, not absent. A human analyst rarely invents a plausible-but-wrong investor out of a fuzzy string; an agent will, confidently, unless you build the guard that stops it. Most of the real work was those guards — the checks that catch a confident error before it becomes a row.

Agents give you leverage on doing. They give you nothing for free on being right. That gap is where the operator still lives, and it's a good place to live.

What a zero-employee company feels like to run

It feels like being an editor, not an author. I'm not writing the migration; I'm the one who asks whether the migration should touch production at all, whether the backup is real, whether the number we're about to publish is one I'd stake the firm's credibility on. The agents propose. The judgment — reversible or not, true or merely plausible, ours to decide or someone else's to gate — stays with me.

That's the model I'm proving, from capital to health tech to trade: not that agents replace the operator, but that they collapse everything between intent and execution, and hand the whole freed-up day back to the only thing that was ever scarce.

Which was never the labour. It was the judgment.

Frequently asked questions

Can a company really run on AI agents with no employees?

NUVC runs on 24 agentic teams and zero headcount. The venture-outcome graph described here — schema, sourcing, entity resolution, and a production database write — was built without a human doing the middle of the work. The operator's role shifts to setting intent and holding judgment on the consequential and irreversible calls.

What is the hardest part of building with AI agents?

Not the orchestration — it's being right. Agents will confidently produce a plausible-but-wrong result (a false entity match, an input mislabelled as an outcome) unless you build guards that catch the error before it becomes a row. The verification discipline, not the automation, is the real work.

Does AI replace engineers or analysts?

In this model it collapses the distance between intent and execution, but it doesn't remove the judgment. Someone still decides whether a change should touch production, whether the backup is real, and whether a number is trustworthy enough to publish. The scarce thing was never the labour — it was the judgment.