Agentic Severance

I spent some time this weekend standing up Garry Tan's GBrain on my personal stack. Postgres with pgvector, hybrid search, the compiled-truth-plus-timeline pattern for every page. The kind of build where you remember why you got into this in the first place.

Then Monday morning hit and I watched the severance happen in real time.

My agents know I was at Richmond Rod and Gun a few Saturdays ago. They know which flight plan I filed out of San Carlos. They know my younger daughter's cheer schedule, which fermentation batch is on day three, and which book I was reading at 11pm. They can reason across all of it.

They have no idea what I did at work.

Not because the data does not exist. It exists. It sits in a Microsoft 365 tenant with Purview labels, a data classification regime, and a GRC posture that does exactly what it is designed to do. The controls are not theatre. I helped argue for some of them. They are the reason our PE clients trust us with their data room, and I would not have it any other way.

But the second brain I am building is only getting half of me.

The Apple TV reference is not just a hook

In Severance, Lumon's employees consent to a procedure that splits their memory cleanly at the elevator. The innie knows only work. The outie knows only home. Neither can consult the other, and the show spends three seasons interrogating what that does to a person.

I did not consent to this procedure. Nobody offered me a waiver. The severance emerged from the intersection of two things that are individually correct. My employer's data controls are correct. My desire to run agents against my own life is correct. The gap between them is where I live now, and the gap is getting wider every week as the personal side gets richer.

The personal agent reads my email, parses my calendar, ingests my chat threads, watches my GitHub repo activities, and writes compiled truth on top of an append-only timeline. It knows that last Thursday I spent ninety minutes debugging a Hermes Agent routing issue and then smoked a tri tip that came out at 196F core. It cannot see that on the same Thursday I reviewed a diligence deliverable that would tell it something real about how I actually think.

Half the signal is missing. The synthesis is partial by construction.

What the severance actually costs

A second brain without the work half is not just incomplete. It is systematically biased toward the parts of my life that have the least structured reasoning in them.

Work is where I apply the most deliberate analytical effort. Where I stress-test arguments with people who are paid to disagree with me. Where I write the longest-form, highest-stakes documents of my week. Strip that out of the training data for my own memory system and what you get is an agent that knows my hobbies better than my craft. It can tell you what I cook. It cannot tell you how I reason about a PE investment thesis, because the reasoning lives in documents it cannot read.

This is not a small omission. For a working professional, it is the omission. The agent ends up being useful for logistics and lifestyle, which is a real but narrow slice of the value proposition. The thing I actually wanted, an agent that could be a thinking partner across the whole of my professional life, requires the severed half.

Where the counter-argument holds

The case for keeping work data out of personal agent infrastructure is strong, and I will not pretend otherwise.

Client data is not mine. Even the patterns of how I touch client data are not mine. Our PE clients have their own data protection obligations flowing through to us, and those obligations do not evaporate because I have a clever idea about memory architecture. If a personal agent ingests a calendar entry titled "TPG portfolio review" and later surfaces it in some benign-looking summary, the leak has already happened. The fact that the agent lives on my DGX Spark at home does not change the classification of the source.

The surface area also matters. A personal agent talks to a dozen SaaS APIs, half of which I did not personally audit. Every one of those integrations is a potential exfiltration path for data that is not mine to put at risk. Our InfoSec team is not being paranoid when they block this. They are being correct.

So I hold both things at once. The severance is the right answer given today's tooling, and the severance is costing me something real.

The Robot Whisperers distinction that matters

The Robot Whisperers position is that humans must not outsource judgment to machines. Re-integration is not an ask to outsource judgment. It is an ask to stop outsourcing memory to the silos my tools happen to sit in.

Memory and judgment are different operations. When I want my agent to remember that I argued a particular position on agentic coding last month and the argument did not hold up by Wednesday, that is not asking the machine to decide anything. That is asking the machine to do what every analog knowledge worker has always done with a notebook, plus search. The decision still lives with me. The agent just needs to have been in the room, or at least have read the meeting notes, to be useful in the next room.

The severance problem is not a judgment problem. It is a context problem wearing a privacy cloak.

What actually has to happen

Re-integration will not come from relaxing the controls. Our PE clients would be right to walk if we did. It will come from building personal-agent infrastructure that the enterprise can actually trust.

That means data residency guarantees that keep work-derived embeddings inside the tenant. It means classification-aware ingestion where a highly confidential calendar entry never leaves the boundary but a public one can. It means memory architectures that can hold a work fact and a personal fact in the same graph without leaking either direction. Microsoft is circling this with Copilot's work-and-web split, but the integration is shallow and the agent sophistication is not there yet. The interesting work is going to happen in the infrastructure layer, not the chat layer.

Until that infrastructure exists and passes an honest GRC review, the severance stays. I will keep running my personal second brain at home and keep the work reasoning in the tenant where it belongs. The two halves do not speak. I will do the synthesis by hand, which is what humans did before any of this existed, and I will feel the friction every time I do it.

If you are building in this space, the opportunity is not another personal productivity wrapper. The opportunity is the boring, unsexy, enterprise-grade plumbing that lets a professional's work life and personal life finally consult each other without anyone's compliance officer having a heart attack. Whoever builds that cleanly gets the severed professionals as a customer segment, and there are more of us than you think.

I would like my innie and outie to meet. Not to share a body. Just to compare notes.