mila-z21
private pre-read

This walkthrough is shared privately ahead of a meeting. Enter the password to continue.

Incorrect password — try again.
to present
agent cofounder skills · demo build for mila ventures studio
mila-z21

An operating system for agent cofounders.

Built at light speed with coding agents for 18 months — long enough to learn that speed was never the constraint. mila-z21 is the discipline I wish we'd run from iteration one, encoded as six skills.

shape-up × agents lacuna signal layer claude · codex · hermes
how the thesis got here · 18 months

I raised $1.25M to build fast.
Speed was never the constraint.

At Alpaca Network we had a small team and coding agents right as they got genuinely good — Claude Code really accelerating around the Opus 4 era. We built at light speed. It still wasn't enough, and that's the whole point.

JAN 2025
$1.25M, as CTO
Raised the seed. Small but mighty team, betting the company on code-gen leverage.
2025 · BUILD
Light-speed execution
Agents hit their stride around Opus 4. We shipped and pivoted faster than headcount should allow.
+3 MONTHS
40k users
Real acquisition, fast. Proof the team could execute on whatever we pointed it at.
THE CATCH
Wrong traction
Users arrived; durable pull didn't. The early framing had quietly set the ceiling.
THE PIVOT
B2C → B2B
We changed the model under load. Expensive — because the path had already compounded.
$1.25M
raised · CTO @ Alpaca Network
40k
users in 3 months
B2C→B2B
the pivot velocity couldn't prevent
the core thesis · drag the launch angle ●

A few degrees at iteration 1 is a different company by iteration 40.

When agents make building nearly free, the expensive mistakes move upstream — into the first few iterations. Path dependency compounds silently: nudge the launch angle a little, and where you land swings several times as far.

iter 10 iter 20 iter 30 iter 40

drag the ● — a tiny angle now, a different company by iteration 40

the thesis, evolved · 18 months

What I stopped believing.
What I bet on now.

Jan 2025 · the bet I made
Build fast. Iterate to traction.

Coding agents make execution nearly free — so ship, measure, and pivot until something sticks.

True, mechanically. 40k users in 3 months. Still not enough.
the first few degrees compound
Now · the thesis I bet on
Execution was never the constraint. Direction is.

When building is free, the expensive mistakes move upstream — into the first few iterations. De-risk the path before you commit the team.

Not "ship faster." Choose the angle on purpose, then commit.

Which is exactly what a venture studio exists to do — de-risk the direction before an EIR hour is spent.

the problem, now generalized · not just mine

Agents run in circles faster.
Theses queue behind people.

agent failure modes

  • Scaffolds unshaped ideas — repo in minutes, direction never questioned
  • 40-task plans on day one — imagined work, not discovered work
  • Builds horizontally — layers and polish before the mechanism exists
  • Progress theater — file counts and checked boxes that figured nothing out

studio bottleneck

  • Thesis-rich, team-poor — the gap analysis lists ventures; testing one needs an EIR + founders
  • De-risking priced in human hours — expensive, serialized, most rows sit unvalidated
  • Betting on slides — which thesis earns a founder search? Judgment, not evidence

Same cure for both: fixed appetites, shaped bets, vertical slices, honest status — Shape Up's constraints, enforced by the agent itself.

three layers · click each

Signal → Process → Evidence

De-risking as a pipeline stage, not a labor cost — it runs before the humans are committed.

signal · lacuna

Where the gaps are

An open commons of ML research: ~753K papers, 38K directions, 191K authors, 3K open proposals. Machine-readable by design.

process · mila-z21

Gap → shaped bet

Six skills run the loop: shape, bet, build, track, hammer, cool down. The agent holds the line at every stage.

evidence · dossier

What the Studio bets on

Per thesis: a one-page pitch, a bet record, a hill chart, and a working slice of the riskiest mechanism.

the operating loop · click a stage

Shape → Bet → Build → Cool down

shape bet build cooldown hill hammer raw ideas enter here — nothing skips the line
status you can't fake · drag the dots

Uphill: figuring it out.
Downhill: just execution.

Agents can fake percent-complete at machine speed — files, commits, checked boxes. The hill asks the only honest question: has anything been figured out?

UPHILL · unknowns DOWNHILL · execution nothing left to figure out

↑ drag a dot — watch the status language change

fixed time, variable scope · toggle the ~

The deadline is the anvil.

Appetite is capped at 8 units and the wishlist weighs 12. Tag nice-to-haves with ~ until the must-haves fit — or the circuit breaker kills the whole bet.

appetite · 8 units fixed

cutting scope isn't lowering quality —
it's how the thing ships whole

not invented — encoded · the venture canon

Standing on the discipline that already works.

mila-z21 doesn't reinvent venture-building. It takes the parts of the canon that de-risk direction and makes a coding agent actually follow them under pressure.

Shape Up · Basecamp

Appetite over estimates

Fixed time, variable scope. Shaped bets, no backlogs, a circuit breaker at the deadline.

→ the spine of all six skills
The Idea Maze

The first turns decide the maze

A venture is a path through a maze of choices; the early turns foreclose the most doors. Path dependency, named.

→ mila-z21-shaping
Product discovery · Cagan

Four risks, hunted early

Value, usability, feasibility, viability — surfaced before the build, not discovered after launch.

→ shaping's rabbit-hole hunt
Working Backwards · Amazon

Write the pitch before the code

Start from the outcome and argue back. The one-page pitch is the artifact you commit to.

→ the shaped pitch + bet
the output artifact · one per thesis

Every gap-analysis row becomes a dossier.

Evidence the Studio's betting table can act on — before an EIR hour is spent, before a founder search begins.

thesis dossier shaped · bet · sliced pitches/example-thesis.md
pitch
Problem · Appetite · Solution
Rabbit holes patched, no-gos declared. Literature-grounded via the Lacuna direction it came from.
bet record
Five questions, one call
Problem matters? Appetite honest? Solution attractive? Right time? Capacity there? — with the agent's own position on record.
hill chart
Scopes + positions
What's solved vs. what's still unknown, riskiest-first — the real state of the technical risk.
working slice
The riskiest mechanism, running
Deployed, end to end. Not a mock, not a deck — the part most likely to kill the venture, proven or broken.

The Studio's mandate, made queryable: bet where Mila's research gravity meets commercial whitespace.

live demo · ~25 minutes

One thesis, whole loop, real artifacts.

02 min

Input

A row from the gap analysis — or a Lacuna direction dense with Mila authors.

07 min

Shape

The agent interrogates: appetite, breadboard, rabbit holes, no-gos → one-page pitch.

03 min

Bet

Five betting questions, agent takes a position, kick-off written.

10 min

Build

One vertical slice of the riskiest mechanism — live, end to end.

03 min

Close

Hill status, one staged scope-hammer refusal, dossier on screen.

Lacuna supplies the gaps. mila-z21 turns a gap into a shaped bet and a working slice. The dossier tells the Studio which theses deserve humans.