Automate almost anything with AI.

Spend more time doing the stuff you love.

Human in command. Machines in motion.

Notes · 0:30

Hello. I run my companies on AI team members, and tonight I want to show you how that works, in plain words and with real numbers. Everything you will see is live, or was read this week. The point is not the technology. It is the hours it gives back.

The idea

Four building blocks

Everything you automate is made of the same four parts, in this order. Learn them once, and almost any job you do at a computer can be handed over.

AgentsWho does the work. A model from one of the labs, given a name and a job.
SkillsWhat the job is. A prompt you keep, in plain words, that gets better every time.
ToolsWhat it can touch. Your mail, calendar, shop, bank or code, through a plug called MCP.
MemoryWhat it learns. Your preferences and corrections, kept for every job after.

Four building blocks, in order

Notes · 0:45

This is not a chat window you type into. It is a team member you hire: with a job description, the tools to do it, and a memory. Four parts: who does the work, what the job is, what it can touch, and what it learns. One slide on each, then some real examples.

01/Agents

Who does the work

An agent is a model with a job. The models come from a handful of labs, and which one is best changes every month. A good setup can swap the model under a job without rewriting the job.

Pick the job, not the lab.

The trust-me-bro benchmarkScore, out of 100

ClaudeAnthropic92
GeminiGoogle91
GPTOpenAI90
GrokxAI86
Kimi K3Moonshot85
DeepSeekDeepSeek84
QwenAlibaba82
LlamaMeta78

Scores invented. The order changes every few seconds, which is roughly how it feels.

Notes · 0:45

Every lab has a leaderboard it tops. This one is invented, and it moves, because that is what the real ones do from month to month. The point is that the model is the least interesting part. An agent is a model with a name and a job, and the job outlives the model: when a better one comes, you swap it under the same job.

Agents / Off the shelf

A good place to start

Every lab now sells an agent of its own. One sign-in, a chat, and it can act in your apps. Try one. Then know two things before you build your week on it.

Good to start with. Hard to build on.

xAI
Grok Bot
Always-on agents, each on its own cloud computer, signed into your tools.
Meta
Muse
A personal agent on your phone and the web, with an email address of its own.
OpenAI
ChatGPT agent
A chat that opens a browser, uses your tools and finishes the task for you.
Anthropic
Claude Cowork
An agent on your computer, working in your files and your apps.

Good to start

One sign-in, no setup, and a real agent that acts in your apps today. Try one this week.

One lab

You get that lab’s model, and only that one. When another lab leads next month, you start again.

xAIMetaOpenAIAnthropic

No central memory

What one agent learns stays in its app. Your preferences, corrections and customers end up in four places that never talk.

The agents the labs sell, and what to know

Notes · 0:45

Every lab now sells an agent. Grok Bot runs on its own computer and signs into your tools; Muse lives on your phone and has an email address of its own; ChatGPT and Claude will open a browser and get on with it. They are a good place to start, and you should try one this week. Two things before you build your week on one. You get one lab’s model, and only that one. And what it learns stays in that app, so your preferences end up in four places that never talk.

02/Skills

A prompt you keep

A skill is a reusable prompt: the job in plain words, kept and run again every day. Every time it gets something wrong you add a line, and the next run is better. No code. If you can brief a new assistant, you can write one.

SkillTidy the inbox

Day 1
Day 3
Week 2
Week 4

Every morning, read what arrived overnight.
Archive the newsletters and the receipts.
Draft a reply to anything that needs one. Never send it.

Day 1

The first brief: three lines and a rule.

Example: one skill over a month. The same prompt, a line longer each time it got something wrong.

Notes · 0:45

A skill is a prompt you keep. Day one it is three lines and a rule. Day three an invoice got archived, so it gets a line. Week two a draft sounded like a bot, so it gets a line about the sign-off. A month in it is still a page of plain words, and it does the job the way you would. That is the whole craft: write it down once, and correct it when it is wrong.

03/Tools

What it can touch

Anything with an MCP server, which is most of the tools you already pay for: your mail and calendar, your documents, your shop, your bank, your code. MCP is the plug. Connect a tool once, and every job can use it.

Google Workspace
Microsoft 365
Slack
WhatsApp
GitHub
HubSpot
Notion
QuickBooks
Salesforce
Shopify
Stripe
Xero
Airtable
Asana
Calendly
Canva
Dropbox
Figma
Intercom
Jira
Linear
Mailchimp
PayPal
Trello
WordPress
Zapier
Zoom
Klaviyo
Miro
Sentry
Todoist
Webflow

Bigger bubbles: the tools most people connect first. Any system with an MCP server fits.

Notes · 1:00

MCP, the Model Context Protocol, is the plug that lets an AI use a system the way you would, with your own sign-in. The big bubbles are the tools most people connect first: their mail, their calendar, their documents. The small ones are the long tail, and anything with an MCP server fits. You connect a tool once, and the model never sees your password.

04/Memory

Stronger every day

One team learns it. Every team knows it.

MemoryEnterprise

Every teamEvery kind81 memories

Mentioned

2026-02-022026-03-24

Source facts

few many

Drag to move⌘ or Ctrl + scroll to zoomPoint at a memory to trace it

Similar meaning, not proof
Close in time
Same entity
Explains

CorrectionIn use

Offer one month free to anyone who asks to cancel.

Replaces ~~Offer 20% off to anyone who asks to cancel.~~

Evidence: You, training it in chat: a month free keeps more customers than 20% off.
Recorded: Mar 20
Labels: SupportAnswer a billing question

Same task: Annual plans renew on the 1st of the month.
Same team: Enterprise customers get a named contact.

AskTrainCorrect

Example: Northstar memory

Notes · 1:00

This is the part that compounds. Every correction you make is kept, with its evidence and who said it, and every job after reads it first. On day one it knows your brief. By day ninety it knows your voice, your product and your customers, so your part shrinks. Remembering is not permission: memory never widens a tool or skips a gate.

Example / My inbox

An assistant for my inbox

One tool connected, ten small jobs: sort what arrives, archive what I approve, draft the replies, and learn from what I actually send. Set up in an evening.

My inbox

One AI member, one tool, one evening to set up

ToolsGoogle Mail

ModelsClaudeGPTGrok

Emails read: 5,930; Threads opened
Archived: 382
Asked me to decide: 31
Lessons kept: 20; In memory
Drafts written: 1; Since 15 Oct
Runs so far: 1,054; Its ten jobs, to 15 Oct

The jobs

Scan and triage personal email41 runs
Review inbox cleanup candidates505 runs
Archive approved inbox messages232 runs
Resolve personal email decisions238 runs
Prepare, check and save reply drafts20 runs
Learn from sent replies and corrections6 runs
Learn from historical sent email9 runs
Review email quality and memory1 runs
Evaluate email replies without mailbox access1 runs
Initialize personal email controls1 runs

My own enterprise, read 15 Oct 2026. The running light is a demonstration.

Notes · 0:45

This is my own mailbox, not a demo. One tool, one AI member, ten small jobs. In the first week it read nearly six thousand threads, archived a few hundred, and asked me to decide thirty-one times. The jobs that archive and draft only act on what I approved, and the learning jobs read what I sent, so the drafts get closer to how I write. It runs every morning before I am up.

What I’m working on

The operating system for your AI-native company.

The four blocks, built as a company you run: AI team members with jobs, the tools they share, one memory, and you at the gates.

Block 01Multi-agentA team of AI members, not one assistant: each with a name, a role and an inbox.
Its ownClonable teams, roles, tasksBuild a team once, or clone one another business shares, with its roles and its tasks.
Block 02Chainable skills (tasks)A task hands off to the next: triage, then build, then review, then ask you.
Block 03One-click tool connectionsPress Connect and sign in. Your mail, calendar, code and payments, shared by every job.
Block 04Centralised memoryOne memory for the whole company. What one job learns, every job knows.

What Zero Human is, in five points

Notes · 1:15

So that is the theory. What I am working on is the four blocks built as a company you run: not one assistant but a team of AI members, each with a job. Teams, roles and tasks you build once, or clone from someone who already did. Skills that hand off to each other, and tools you connect once, with a click, shared by every job. One memory for the whole company. The next slide is it running.

Zero Human / Live

The team, right now

Every run Zero Human’s own team starts and finishes, as it happens. Nothing here is staged: it is the company that builds the product, building the product, while we talk.

Notes · 1:30

This is live. On the left, the team: who is working and on what. On the right, the runs as they start and finish, read every ten seconds. Watch it for a moment: an issue triaged, a pull request reviewed, a test suite run. Nobody is driving this. The founder set the goal and holds the gates, and the team fills the day.

Our ask

Go automate something

Pick one job you do every week and would rather not. Sign up, connect one tool, describe the job in plain words, and let it run for a week. Then tell us how it went: what you automated, and how well it worked.

QR code for zerohuman.com/curiosity — Where to go next

Notes · 1:00

Here is the ask. Go to zerohuman dot com slash curiosity. The full instructions are on the page, with a free plan and the three steps on this slide. Pick one boring job, hand it over, and tell us how it went, good or bad. We read every reply. The link stays up while we take questions.