MPS All-Hands

Is AGI here today?

Humanity’s Last Exam · 2,500 questions like this

Could you answer this?

A quantum-information question
“Two qubits share the state |ψ> = ( |00> + |01> + |11> ) / √3. What is the von Neumann entropy of the first qubit’s reduced state, in bits?”
Answer with a number.
Experts average ~90% — only in their own field. Nobody has the range.
The climb

The best generalist alive.

0255075100% A specialist, in one field: 90% 8%16% 35%53% Jan ’25Aug ’25early ’26Jun ’26
The problem

So why doesn’t it feel like it?

The gaps — what it’s missing
Isolated apps
walled off from each other
No persistent memory
forgets every session
No real context
blind to your day
Not always-on
only when you summon it
The model is brilliant. The assistant in your pocket still isn’t.
What I believe

Everyone gets a butler.

Once, only for the rich
A driver
An executive assistant — the kind your VP has
A personal nutritionist
A lawyer on call
Now: an agent.
For everyone.
Once you’ve had one, you never go back.
Where this started

OpenClaw did what the big labs didn't.

KClaw
the test machine
evolves into
Eevee app icon

Eevee.

an agent that evolves to fit you.

Feb 2026
OpenClaw → Galaxy
Early Mar
First demo to Sally
Late Mar
Native Android app
Apr–Jun
Mini-apps · connectors
July 2026
Dogfood · ESIL × AIL
Under the hood

How Eevee actually works.

Eevee / KClaw architecture: entry points and gateway feed the agent runtime and a persistent agent core (memory, identity, reasoning, planning) drawing on context streams and device sensors; every action passes the Trust / Policy Gate before reaching device APIs, alongside the cloud model layer, on-device memory, audit log and background workers; dark features off by default
Why you'll want one

Ask once. It's handled.

Persistent memory
Your context, recalled every turn.
Long-horizon tasks
Runs for hours — replans when blocked.
Self-improving
Nightly reflection — it learns your patterns.
Tool use
Real actions in your apps — new tools every week.
Your phone knows your life. Finally it acts on it — and hands your time back.
Demo · 01 — recorded on a real Galaxy

A study plan that knows its student.

The ask
“Help me create a study plan for the last weekend of July — and schedule it on my calendar.”
I never said the subject — it knew: corporate finance, my MBA exam
Built for my gap — years in tech, zero finance
Checks my calendar first — then books the plan in
No intake form, no syllabus hunting. It already knows its student.
Demo recording
Demo · 02 — recorded on a real Galaxy

Rambling in. Clean notes out.

You, mid-YouTube
“Eevee, jot this down — that retention thing he just said, uh, we could try it on onboarding, something like…”
It sees what you see — catch the idea without leaving the video
Filed into your knowledge base — a dedicated space
Memory brings it back when it matters again
Dump ideas anytime. The right one resurfaces at the right moment.
Demo recording
Demo · 03 — recorded on a real Galaxy

Build a routine. Eevee runs it.

The standing job — free charging ends at hour four
“When my car starts charging, remind me to move it before the free hours run out — and work around my meetings.”
Charging notification → Eevee knows the clock started
Reminder on your calendar, before hour four
In a meeting then? It moves the reminder earlier.
No tracking, no mental math, nothing to remember. That’s what a butler is for.
Mini-app creation

It doesn't have to be useful.

The case against keeping it
Expensive — every build burns real tokens.
Doesn't even touch the vibe-coding stack.
Months of hunting — no killer useful use case.
Until I built one myself with KClaw’s mini-app creationInner Peace: swipe away “per my last email” to reach Enlightened (Resigned).
Why does everything have to be useful? Sometimes you just want to have fun — and express yourself.
The Eevee dogfood

Test 100 ideas in days — not months.

Idea
Vibe-code it
Ship to dogfood
Learn from use
↺  into agent design
What we provide
$ally covers the tokens
The app — just an APK
Weekly build workshops
We ship your fixes fast
What we need from you
Your own Galaxy
Broad permissions
A real use case
Daily use, shared openly
Cross-lab · ESIL × AIL~10 → ~30 people6 weeksJuly 2026
Bring the one task that annoys you — we'll build it this week. → #eevee-dogfood-voc
Before & during dogfood

How we’ll know it’s working.

Automated evals
LLM-judged test suites — the unit tests of an agent.
Basic functionality, checked on every build.
Vibe checks
We eat our own dogfood — the team is customer 0, on Eevee daily.
No metric replaces living with it.
Community feedback
Explicit: report issues, pitch features — in-app, workshops, #eevee-dogfood-voc.
Implicit: session data, LLM-analyzed for what works — pipeline ready, light touch for now.
Autotests catch regressions. Vibe checks catch weirdness. You find the use cases.
The shift

Where this all goes.

Today vs Future: today's app-centric Android stack (apps, platform layer, Android OS, hardware; fragmented context) versus a future agent-centric ecosystem with You and Your Agent at the center, always present across phone, watch, earbuds, car, laptop and home display, drawing on persistent memory and identity, device signals and context, tools and actions, and an open agent ecosystem of agent-to-agent collaboration, MCP services and generated apps
This is just the beginning

Let’s find out — together.

Stay curious
Necessary chaos
Learn fast

Dogfood is more than a test machine for use cases — it’s how we build a culture of curiosity, and learn faster than anyone.

Bring a Galaxy and your wildest ideas — #eevee-dogfood-voc
1 / 13
EDIT MODE · click any text · Ctrl/⌘+S to save · Esc to exit