The next class of AI — agents that stand in for you, not assistants that wait
Aardvark learns how you write and how you decide, then handles the work in your voice for the whole stretch you're gone — overnight, mid-meeting, or a week into a trip with your phone in a drawer. When something lands, it drafts the reply, the memo, the redline the way you would have, and holds it. You come back to finished work, cited and ready. Nothing leaves until you click approve. It's an early, honest step toward a bigger shift: agents that take your function and stand in for you — newly possible because models can finally write in one person's voice from their own priors.
Who it's for
A private beta, one ring of individual users at a time — we start with the people whose writing is unmistakably their own, the hardest and highest-trust first users. Everything runs inside your own boundary — no shared accounts, no fabricated names here.
The frontier it's built toward
Not an assistant that waits for a prompt — an agent that holds one piece of your work and does it in your name while you're elsewhere. Aardvark doesn't claim to be all the way there: it reads across your work, drafts in your voice, and holds — and you approve. It's an honest first step, and it earns the claim four ways.
Security you can see
Inside your own boundary, encrypted at rest, read-only by default. Every action is written to an audit trail — and nothing is applied or sent without your click.
It reads what you'd have to
Before it writes a word, it reads across the whole request — the thread, the files it points to, their attachments, and your own prior work — in whatever format they arrive: documents, spreadsheets, slides, PDFs. Then it connects what's related across those files: a clause in one attachment that changes the reply, a number in a sheet that contradicts a memo, a commitment buried in an old thread. It surfaces what matters and pulls the precedent to draft from — and cites every claim back to its source. It doesn't know everything; it reads across what you gave it, so the connections buried across your files surface — instead of you being the one who has to read every attachment to catch them.
Your voice, on the page
It studies your own past letters, replies and edits to learn how you write and how you decide, then drafts what you'd likely have written — for you to review. Every line carries provenance — hover any clause to see where it came from. And it keeps learning: every draft you approve or set aside is training data only you have — a corpus of one, your judgment learned from your own priors and never pooled. That's exactly why no shared foundation model can reproduce it: the training signal never leaves your boundary. It gets sharper, and harder to leave, the more you use it.
Draft · hover to trace each clause
Happy to move the timelinehow you replied in March · 41 prior notes to the 14thyour calendar · next open slot, and I'll send the revised scope before thenthe thread above · their ask.
It works while you don't
You don't queue it up or check in. Work lands — 2am, mid-flight, mid-lunch — and Aardvark is already on it. You come back to a finished draft, not a to-do — reviewed by you, never auto-sent.
One stretch away, start to finish
Here's one real run — an afternoon in back-to-back meetings, though it works the same overnight or a country away. Every step logged, inside your boundary, and stopping short of the one thing only you should do.
A request comes in while you're heads-down — a reply owed, a memo, a redline. Aardvark picks it up. You're away, and you stay away.
Read-only, it takes in the thread, the two attachments and your own prior notes — then connects them, and catches the clause in the Q2 addendum that changes what the reply owes. The thing you'd only have found by reading everything.
A full draft in your voice, built from your own precedent, every claim traced to its source — the Q2 clause among them. Not a summary — the actual work, ready to go the moment you say so.
Out of back-to-backs, you return to finished, cited work and a receipt of exactly what it did. You review, you edit, you approve. It never sent a thing. The gate never moves — nothing is ever sent without you. What grows is how much you let it carry: you start with one kind of work, and as it proves itself in your voice, you hand it the next.
The honest receipt
Every run leaves a replayable record — what it read, what it wrote, and what it deliberately did not do. Read-only by default; sending is a human act; your documents stay inside your boundary. This is the receipt from the draft above, exactly as it was written.
In your boundary, not ours
Runs against the files where they already live. Your documents are never handed to a third party or used to train anyone else's model — they stay within your own boundary.
Read-only by default
It reads to learn and to draft. Writing to a file or sending a message is a separate, explicit, human-approved step.
Every action audited
A durable, replayable trail of each read and each draft — timestamped, attributable, and yours to inspect at any hour.
Nothing sent without you
The hard line. Aardvark prepares; a human approves and sends. There is no path where it sends on your behalf. The gate is where trust is earned, not where it stops — as it proves itself on one kind of work, you let it carry more of that work, always your call, one class at a time.
Invite-only beta
We're opening a small first ring — people whose writing is unmistakably theirs, who won't hand that to a generic chatbot. If your name on a document carries weight, we'd like to talk.
The first ring shapes what standing in for someone should mean — we'd rather learn that with them than declare it.