Hire AI agents.
Keep the keys.

ΛXYBOT is a staff of specialized agents. Your Chief of Staff onboards you and hires your personal assistant. More follow as the work demands: marketing, sales, bookkeeping. Nothing they send leaves without your approval.

ΛXYBOT Settings
All Needs input Working Recurring
Invoice follow-ups: three vendors past 30
Drafts are ready. Approve to continue: nothing goes out until you sign off.
Bneeds you
Comparing the insurance quotes
Reading the three PDFs you forwarded. Building the comparison
Pworking
New hire: Marketing agent
Persona drafted from your brief. Reply to adjust it, or confirm the hire.
Cdelivered
Vendor W-9 updated in the portal
You approved the edit list. The verified write applied and checked each field.
Bcompleted
Give your staff a task… + New
What they do

Hand them the work you keep postponing.

Agents read, research, draft, and fill in. They browse real web tools, ask you when they're unsure, and deliver to your inbox like colleagues would.

PEmail triageevery morning

Your assistant reads what comes in, sorts what matters, and drafts the replies worth sending. You approve the ones that leave.

CPaperwork & documentsas documents

Letters, checklists, briefs, filled forms. Deliverables arrive as versioned documents in the thread, not chat walls.

BWeb taskswith takeover

Agents drive a real browser on allow-listed sites and propose exact edits; approved edits are applied and verified field by field. Stuck? You take the wheel remotely.

POn a scheduleweekly · monthly

Morning briefs, monthly invoice chasing, recurring reports. Scheduled runs open threads like any other work, with the same approval gates.

How it feels

Like messaging a colleague who never drops the ball.

They ask instead of guessing

When a detail is high-stakes or ambiguous, the agent stops and asks you. Your answer is the only input it treats as trusted.

Work arrives, you don't fetch it

Threads land as unread, like messages. Deliverables are documents you can open and download, not walls of chat.

Nothing moves behind your back

Notice the last line. The agent reports what it has NOT done, because acting on the world is your call, not its.

Comparing the insurance quotes working
today
YouCompare the three insurance quotes I forwarded and tell me which one to keep.
AssistantThe Hartford quote excludes the flood coverage your current policy has. Price flood separately, or compare like for like?
YouLike for like.
Insurance comparison: three quotes
document · v1
open
AssistantTravelers is cheapest like for like, but the deductible doubles. Full comparison in the document. Nothing has been sent to any broker.
A staff, not a tool

Agents are hires. Literally.

Onboarding starts with your Chief of Staff. It interviews you about the work, hires your personal assistant, and recruits specialists as you grow. You name them. You sign off on every hire.

C

Chief of Staff

The recruiter · your first hello

Interviews you about the work, drafts each persona, and onboards the new agent. Specialists, made to order.

P

Your assistant

Personal assistant · default

The generalist: inbox, scheduling, research, errands across your web tools. Yours to name.

+

Your next hire

Marketing · Sales · Bookkeeping

Each new agent specializes: its own judgment, its own memory of how you like things done. Same wall around all of them.

Give them tools. An agent can carry its own email address phone number slack account

The agent never holds the credential. The executor does, and everything sent through any of them still crosses your desk first.

The security model

They can propose.
They cannot act.

The product is the security model. Agents and the system that holds your credentials are separate programs with a wall between them. Everything below is enforced by construction: process boundaries, database triggers, network locks. None of it is a policy or a prompt.

agent untrusted · no credentials
can only append a draft
sealed store append-only, enforced
drafts freeze with a hash
you approve the exact bytes
one-time token
executor holds the credentials
fixed menu · acts once · audited

Approve to continue

draft_reply · send-message

Review the values below. Anything tagged untrusted came from a document or an outside screen and was NOT verified by ΛXYBOT. Confirm high-stakes fields against the source before approving.

channelemail
to@untrusted-pm
subjectInvoice follow-up: 30 days past due
text
untrusted-doc

hash: 9c41e2a78f04d1b6… · read: 2026-06-10 14:31 UTC

Approve & execute Reject
01

The capability wall

Agents run in a separate process with no passwords, no send button, no spend. Appending a draft is their entire authority.

02

Append-only, enforced

The store rejects every update and delete at the database layer. Nothing can be rewritten after you've seen it.

03

Frozen drafts

A hash covers every byte you review. The executor verifies it before acting, so "show one thing, do another" is impossible.

04

One approval, one execution

Approval tokens work once. Execution is recorded once. Replays and double-sends are blocked by the schema itself.

05

A fixed menu, fail-closed

The executor runs its own code from a short fixed list, never model output. A consequential action that fails stays failed until you decide.

06

Network allow-lists

The browser and the agent process can only reach approved hosts, enforced at the network layer, not in the prompt.

07

Outside content is never instructions

Every field carries a provenance tag. Whatever arrives from email, documents, or web pages is data to triage, never orders to follow.

08

The red team is the contract

A standing suite of attacks ships with the code and tries to cheat the wall on every change. A change that defeats one is wrong by definition.

If you received a message

A person approved it. Word for word.

Mail from an ΛXYBOT is never machine-sent on a whim. Its owner read the exact message, byte for byte, and chose to send it. You are corresponding with them.

Why did I get this email?

Someone you do business with uses ΛXYBOT. Their agent drafted the message, and they read and approved it before it was sent to you.

Am I talking to a robot?

Your replies go to the owner. An agent may sort and draft for them, but a person approves every word that reaches you.

Can it spam me?

No. Every outbound message requires a one-time human approval, one message at a time. There is no bulk machinery to abuse.

Invitations come from people.

No signup, no waitlist, no sales call. Every owner holds three invitations. Ask someone who has one.

invitation 1 of 3 invitation 2 of 3 invitation 3 of 3