Run z21 on the Studio's gap analysis.
z21 is a set of agent skills. Point them at a research gap and they return a shaped bet and a working slice. 18 months of building with coding agents, distilled into six skills and a 25 minute demo.
mila-z21 doesn't reinvent venture-building. It borrows the proven parts of de-risking direction and gets a coding agent to actually use them.
Mila Ventures Studio turns AI research into companies as an operational co-founder for a small, founder-friendly stake: a dedicated EIR from inception, product to MVP, first customers via 100+ partners, co-founder matching, IP and fundraising support. It de-risks the business from lab to adoption.
Fixed time, variable scope. The cap sits on the cycle, not on your ideas. A circuit breaker at the deadline.
A venture is a path through a maze of choices; the early turns foreclose the most doors. Path dependency, named.
Value, usability, feasibility, viability: surfaced before the build, not discovered after launch.
Start from the outcome and argue back. The one-page pitch is the artifact you commit to.
At my last startup — small team, coding agents right as they got good — we raised on a bet and spent 18 months in the idea maze. At the fork that mattered we ran both ways instead of picking one, and two paths spun out clean. In the end, the early turns mattered more than anything we built.
dead ends everywhere, one way through — except the fork, where we ran both ways at once.
When agents make building nearly free, the expensive mistakes move upstream, into the first few iterations. Nudge the launch angle a little, and forty iterations later you land somewhere very different.
drag the arrow: a tiny angle now, a different company by iteration 40
Build fast, ship, measure, pivot until something sticks. Coding agents make execution nearly free, so lean on speed.
When building is free, the expensive mistakes move upstream into the first few iterations. Which angle you pick there is what compounds.
Which is what a venture studio is for: de-risk the direction before an EIR hour is spent.
Big bets need conviction to get funded. The usual way to earn it — slow, serial de-risking by committee — drains the ambition that made them moonshots.
You can't manufacture conviction — but you can earn it cheaply. Point an agent at a messy direction and get a testable slice back in hours, not months: evidence enough to commit while the spark's still lit.
Agent skills 10x the studio's core work: validate a bet, match a direction to the expert who should own it, and more. De-risking becomes a defined process any domain expert in Mila's network runs, agent-assisted, before an EIR or a founder is committed.
The whole map, machine-readable, so an agent reads it, not an EIR: ~753K papers, 38K directions, 191K authors, 3K open proposals. Author density matches each direction to the experts who could run it.
Six skills run the loop: shape, bet, build, read, cut, debrief. Defined tightly enough to hand a domain expert, agent-run and parallel, not serial human hours.
Per hypothesis: a pitch, a funding record, a signal readout, and a working slice. Fund on evidence, not slides.
Agents can fake percent-complete at machine speed: files, commits, checked boxes. The hill tracks a different thing: has the risky part actually been figured out yet?
↑ drag a dot, watch the status language change
Building is cheap now, so nothing stops scope from sprawling. The cap isn't a labor budget, it's how much this bet is worth proving: 8 units against a 12-unit wishlist. Click an optional row to cut it until the risky must-haves fit one cycle, or the circuit breaker stops the bet.
cut scope to ship it whole,
inside the time box
Evidence the Studio's portfolio review can act on, before an EIR hour is spent, before a founder search begins.
The Studio's mandate, made queryable: where Mila's research strength lines up with open commercial space.
A row from the gap analysis, or a Lacuna direction dense with Mila authors.
The agent interrogates: appetite, breadboard, rabbit holes, no-gos → framed hypothesis.
Five funding questions, agent takes a position, kick-off written.
One vertical slice of the riskiest mechanism, live, end to end.
Signal status, one staged scope-control refusal, dossier on screen.
Lacuna supplies the gaps. mila-z21 is the defined process a domain expert runs to turn a gap into a framed hypothesis and a working slice. The dossier tells the Studio which bets deserve a team.
Bring a shelf of ideas with no cap and no kill criteria. That is a betting table waiting to happen. Install the skills, ask the agent what to work on next, and see which idea earns a cycle.
git clone https://github.com/Uptick-Systems/z21
cp -r z21/skills/mila-z21-* ~/.claude/skills/