The Loop Is the Product

Agents can do the work of a whole team now. Most of us are still running them one prompt at a time.

I lived in that world for a year. Forty tabs, a different AI in each one, a prompt for everything. I would get a great answer, close the window, and the work was gone. Nothing stuck. I was working harder than I ever had and ending each week roughly where I started it.

The thing that changed everything was not a better prompt. It was a loop.

A prompt tells an agent what to do. A loop gives it a job. The difference is that a loop remembers. It does the work, checks the work, ships the work, and then writes down what happened so the next run starts smarter than the last. Once I built that loop, I stopped being the bottleneck and started conducting. That is how I run the whole portfolio now: each venture on a model I tune to it, getting better while I sleep.

This is the first thing I wanted to write down, because everything else I do sits on top of it.

First, a substrate

Before any of the smart stuff, you need a place for the work to live. I call it the substrate. Mine is a vault of plain text files, and which app sits on top of it matters far less than the fact that the files are mine.

It is not a notes app in the way you are picturing. It is the single source of truth for every company I touch. Every proposal, every pitch, every meeting, every decision, every person, every outcome lands there as a plain text file, linked to everything it relates to. It is append-only and queryable. Nothing gets lost because nothing gets thrown away.

Security is built in. When a venture needs it, it gets its own brain, walled off from the rest, kept in its own store rather than one shared folder. We connect the dots on what is reusable, the patterns and the playbooks, never the proprietary data itself. The learnings travel between ventures. The secrets stay home.

The idea I stole comes from Andrej Karpathy. Software used to be lines of code a human wrote. Increasingly it is knowledge a system compiles from examples. So I treat my vault the same way. Raw material goes in (an email thread, a transcript, a contract), and the system compiles it up into a higher layer: concept pages that synthesize many sources into one understanding. The vault is not where I store information. It is where information gets refined into knowledge.

On top of that sits a knowledge graph, rebuilt on a schedule across the whole vault. It reads every link between every file and draws the map: which ideas are central, which clusters are forming, where the gaps are. So I am not just searching my own brain. I can see the shape of it.

That is the power of continuous learning, and it is the unglamorous part. The vault is the floor. Everything compounding happens because there is a floor to compound on.

The ghost at the center

Once you have a substrate, you need a way to stand on top of it instead of drowning in it. That is the ghost. We are building it as Ghost, the cockpit I run the whole fleet from.

The picture I kept coming back to is not a chat window. It is a glowing core in the center and a fleet of agents in orbit around it, each one a little ghost with a job. Most sit dormant. When one wakes up to do real work, it lights up, fires a beam back to the center, and the core watches it run. When it finishes, it writes its result back into the substrate and goes quiet again.

That picture is not a mood board. It is the actual interface I work from. I set the goal, the fleet does the finding, building, checking, and shipping, and I step in only at the few moments that need a human. The ghost is at the core of everything we build because it is the thing that lets one operator run many companies without losing the thread on any of them.

Then, evals. This is where it gets smart

Here is the part worth getting to, and it is the whole reason the work gets better instead of just faster.

Let an agent roam free and you get expensive slop. Confident, fluent, wrong. The fix is to box it in: give it a clear goal and then grade it at every step. That grading is an eval. It is just a rubric and a scorecard, run automatically, every time, against a frozen set of examples I already know the right answer to.

I write a custom eval for every kind of work that matters. For work where reality eventually tells you if it was good, I keep a library of past examples labeled with what actually happened. Proposals that won, proposals that lost. Before I send a new one, the system scores it against the winners and tells me, plainly, whether this looks like one that closes. For work that only a human can judge, like whether it is in my voice or the best of three options, I grade against a written rubric and I log every single time I overrule the pick. That override log is the cheapest, highest-signal training data the whole system produces.

The approvals themselves moved past yes or no. The gate runs four buttons, not two: Reject, Approve, Approve and Benchmark (this one beats the current best, learn from it), and Approve with a Note (I almost killed this because of X, remember that). Two of those four buttons exist only to capture how I think, so the agent gets sharper every cycle, on my taste, not a generic one.

I started with proposals. Then pitch decks, scored on a dozen dimensions like narrative arc, clarity of the ask, and whether the proof actually lands. Then, the one I did not expect to work: video. I do some video work for partners, and that pipeline generates a batch of clips on every run. A clip evaluator scores each one on brand, lighting, framing, motion, and whether it serves its story beat. The good ones become the new bar. The next run has to beat them. So the video compilations are not just getting made. They are getting smarter, run over run, the same way the proposals did.

Every closed loop makes the next output better, and the substrate remembers all of it. That is compound learning.

What this actually is, on my screen today

People hear "brain" and picture a tidy notes folder. What Ghost actually is, today, is a bird’s eye view of everything I am running.

One screen. Your vault is the source of all truth: a master brain in the center, every venture in its own vault orbiting it, every active agent lit up and reporting back. I can see the whole portfolio at a glance, then drop into any one company and see its pipeline, its work in flight, and what is waiting on me.

Every morning it hands me a daily brief. Not a dump of everything that happened. The synthesis: what moved, what needs a decision, what an eval flagged, the three things worth my attention. The weekend’s learning is already folded in by the time Monday starts.

And it runs a board that fills itself. Think of an automated Trello or Jira, except I am not the one dragging cards across it. The system finds the work, opens the card, does the work, moves it, and closes it, surfacing only the gates that need me. Ten minutes with it and I have a crystal clear snapshot of the whole portfolio: what moved, what is stuck, what needs me. The rest, the fleet handles.

This is not a roadmap I am pitching. It is how I work right now. Where a piece is still hardening, I will say so rather than oversell it.

Why this is the first thing I wrote

Because the loop is the product. Everything else I will write here, a teardown of a venture, a proposal that closed, a deck that raised, is downstream of this one idea. It is easy to ship and guess. The unlock is to measure and improve, and the gap between those two compounds every single week.

Resolve has two meanings, and I mean both. Close what is open. And stay resolved in the pursuit of excellence. Build the loop once, and it does the first while you hold the second.

If you have a real opportunity and you are tired of running it one prompt at a time, that is the conversation I want to have. We build the loop on your business, with you, and it is yours to keep.

Build the loop once. It closes what is open while you go find what is next.

A note from later

This one carries a title I would not choose again. The loop is the machine, and the machine is real, and everything described above still runs. But the loop is not the product. What the loop deposits is: the record of what got decided, what evidence backed it, who overruled it, and what happened after. That is the asset, and this piece is how it gets built.

I am leaving the title alone because rewriting my own history to look consistent is exactly the habit this whole site argues against. The correction is The loop was never the point, and the argument for why the record beats the model is Evidence compounds.