Vision

Where this came from

Look at the work at U2V from a distance and almost every task is the same shape. There is data work — structuring messy information, storing it, retrieving it later — and there is decision work — judging what it means and acting on it.

Filtering a list of inbound startups, writing structured records to Attio, pulling a curated list for a potential LP, deciding which startups are worth a first call: each of these is really two jobs stitched together. One part moves and shapes information. The other part makes a call.

That split is the whole idea behind this system.

The bet

I don't believe AI should be making U2V's investment decisions — not during this fund, and not as a near-term ambition. No one makes those calls better than the U2V team, and building a system that tries to would aim the technology at exactly the thing it's worst at.

The value is on the other side of the split. Structuring and retrieving information — quickly, consistently, at scale — is something agents genuinely do better than we do by hand. So that is where AI at U2V should focus, and where this system deliberately draws its line.

The organizing principle

The agents do the data work. The team makes the decisions.

Concretely, that means two teams working side by side:

• The agent team turns messy, unstructured market noise into clean structured knowledge, stores it in Attio, and delivers it back to the humans exactly when it creates the most value.
• The human team spends its time where judgment matters — moving startups through the funnel, matching them to the network, and building relationships with LPs.

The ambition is to let the agent team fully own the data-structuring and processing work, so the human team can concentrate solely on what drives the fund: informed investment decisions and relationships.

The foundation: control the memory

An ambitious system has to start at the base and build up. Here the base is memory — a knowledge hub that holds the structured information the agents write, and that they can draw from whenever a decision needs it. Every decision is powered by this layer.

For U2V that hub is Attio. It is the single source of truth where the fund's proprietary knowledge lives, and it is the fuel for every agentic workflow. Which leads to the conviction the entire architecture rests on:

An agent's work is only as good as the context it stands on. So the data quality in Attio has to be of the highest possible level — because that is what the agents read back when they reason.

Get the brain right, and everything built on top of it gets better for free.

Two types of agents

Once the goal is a high-quality company brain, the agents sort naturally into two kinds — and every automation in this repo is one or the other:

flowchart LR
    Noise["Unstructured<br/>market noise"] --> Writers
    Writers["✍️ Writer agents<br/>structure &amp; store"] --> Attio[("🧠 Attio<br/>the company brain")]
    Attio --> Retrievers["🔎 Retriever agents<br/>surface when useful"]
    Retrievers --> Humans["👥 The human team<br/>makes the decisions"]
    Humans -.->|new knowledge| Writers

• Writers feed structured knowledge into the brain — turning a pitch deck, an LP website, or a one-line request into a clean Attio record. (dealflow, lp-intake, lp-tasks.)
• Retrievers get knowledge out of the brain when it's most useful — a sharing list for an LP, a matched shortlist, a sample deck for the data room. These are strictly read-only. (lp-share, lp-match, dealflow-sample.)

Every agent owns one distinctive job — one precise workflow, start to finish. And the workflows group into the two things U2V actually spends its time on: dealflow and fundraising.

What we've built so far

Over the last weeks — increasingly together with Adriano — each workflow started from a real task that landed on our desks. The guiding question was never "how do I solve this once?" but:

Instead of solving it once, how do I build a system that solves this repeatedly, autonomously, at a fraction of the cost — and long after I'm gone?

That produced the first set of workflows, all reachable through one shared inbox — Klaus (klaus@u2v.vc). When a task recurs, you hand it to Klaus:

1. Process and upload inbound startups to Attio
2. Build the dealflow-sharing Excel list for our LPs
3. Select startups from our dealflow that match a company or a sector
4. Upload a new LP lead to Attio
5. Produce a dealflow-sample document for the data room

This is only the beginning. The architecture is built so that adding the next capability is as cheap as possible — see a repeating task, and it becomes the next candidate to hand off.

Where we stand, and where we're going

Today the foundation and the first workflows exist: several are live and in daily use, the rest are built and in final testing (each automation page notes exactly where it sits). The brain — Attio — is the anchor, and the write/retrieve loop around it is real.

The direction from here is to widen that loop until the agent team fully owns the data work. If we get there, three things follow:

1. More efficiency, lower cost — the repetitive data work runs itself.
2. Reach — 4,000 startups screened per year — a throughput no manual process gets to.
3. Better investment decisions — because the team's time, and a higher-quality brain, are both pointed at the judgment that only humans should make.

Keep reading

The Design principles show how this vision is enforced in code, and the Automations walk through each agent one at a time.