Old Growth.

A studio that builds companies with agents instead of employees.
One company a week. Same playbook, different tree.

How it works · how we make money

The conviction

An employee doesn't have to be a person.

— Pete Galloway · founder · Deep Forest AI

The problem

Building a company takes 100 people. Most of the work is repeatable.

Founders raise money to hire engineers, marketers, ops, sales, support, finance, legal. Then they hire managers to manage those people. Then HR to manage the managers.

Five years and $50 million later, the company that was supposed to ship one product ships … a HR department.

Almost every job inside a company is a function. Functions are agent-shaped.

The math has shifted.

Frontier model inference is now cheap enough to run a customer-support team for less than a single human's salary. Cold-outreach automation has caught up with SDR floors. SEO content generation outpaces five-person content teams.

The pieces are here. We're going to assemble them.

The operating model

Two humans. Everyone else is an agent.

Deep Forest AI — stealth tech parent. Where Atlas's engineering team builds. No public face. The factory floor.

Old Growth — public studio. Marketing, operations, ventures, AR, legal. Where customers meet whatever Deep Forest just shipped.

Pete — founder · final venture call · de facto operator
Tod — advisor · Ventures committee vote · COO seat pending sign-on
Jordan — operator-lens advisor · CEO of Calendar Island Cannabis Co.
Ash — Chief of Staff agent · keeps the studio moving
Atlas · Lee · Hollis · Auden · Marlow — function leads, all agents
100+ specialist agents below the leads

The team

Pete owns the direction. Ash owns the day.

Founder

Pete

Sets the direction. Reviews Sunday's launch packet. Names every tree. Talks to customers. ~25 minutes per launched company.

Chief of Staff

Ash

Pete's right hand. Keeps Pete + Tod + Jordan moving. Drafts emails, routes ideas to Hollis, notifies Pete when something needs his eyes. Sonnet 4.6 with tool access.

Function leads (agents)

Atlas · Lee · Hollis · Auden · Marlow

Engineering · Marketing · Ventures · Agent Resources · Legal. Each runs a team of 4–28 specialist agents. Each has a charter, a voice, a rule they enforce.

All five leads, plus the >100 specialists below them, are documented in the AR registry. None of them have a payroll line.

The intake pipeline

Five signals before Pete decides.

1 · Surface

Hollis

28 industry scouts each own one vertical. One idea per scout per week, ranked by Hollis's score (0–10).

2 · Gate

Marlow

Legal feasibility memo on every idea. GO · Conditional GO · NO-GO. Outside counsel called when needed.

3 · Vote

Committee

Tod (ops lens) + Jordan (operator lens). Advisory only — does this run after launch?

4 · Decide

Pete

Sees all four signals on one card. Approve · Defer · Reject. The only person who can change state.

5 · Build

Atlas

Approved Monday → live Monday. The seven-day cycle starts.

The build cycle

Pete's idea on Monday. Public launch the next Monday.

Mon

Pete

Approves idea in Ventures. Atlas takes over.

Tue

df-architect-01

Ships the spec. Picks library reuses. Flags net-new.

Wed

Parallel build

codegen · agent-assembler · infra · payments · domain. All running.

Thu

Parallel build

Code freeze EOD. Deployment to staging.

Fri

Integration

All agents wired together. Staging hot.

Sat

df-qa-01

tuesday-rehearsal validation suite. ~30 checks. Pass required.

Sun → Mon

Pete · Tod

Sunday review packet. Monday 09:00 ET launch + handoff.

Pete's total time per launched company: ~25 minutes.

The moat · two compounding assets

The visible compounding is the library. The hidden one is the data.

Visible · the library

If a pattern shows up in two trees, it goes in the library. The third tree gets it free. Build № 6 takes three days, not seven.

Pattern

From

Reused

tpl/stripe-webhook

Yew

every paid tree

tpl/tuesday-rehearsal

Sorb

every launch QA

tpl/customer-support

Yew

support-heavy trees

tpl/meta/cost-optimizer

Pine

every future tree

Hidden · per-vertical data

The library is copyable. Anyone with the same model access can rebuild it in a quarter. The harder asset is what each tree's agents learn about their specific vertical after 90 days of operation.

Birch's claim agents, after a quarter in production, will know how ~4,000 specific payer rules quirk. Pine's scrapers will know which heating-oil dealers in Maine post stale prices on Tuesdays. Sorb's mitigation agents will know which DDoS patterns each customer's edge cluster sees.

That data is the moat. The library is the head-start.

The portfolio · what's already in flight

Three trees up. One approved. Forty more to come.

● Live · pre-volume

Taxus baccata

Yew Privacy

Privacy-as-a-service for car dealers. GDPR/CCPA review at scale. Workers + D1 + Email Routing on Cloudflare. 8-agent fleet.

● Just launched

Sorbus aucuparia

Sorbnode

Edge node monitoring + auto-mitigation. 23-component fleet. Tuesday commercial launch, first customers acquiring.

Launching Friday

Pinus strobus

Pine Prices

Pricing data + outlier detection for SMB e-commerce. 29-agent fleet, code-complete, gated only on Marlow's scraping memo.

⚙ Approved

Betula papyrifera

Birch

Independent-practice billing — solo MDs recover 10–12% of revenue lost to billing errors. Hollis 8.2 · Marlow Conditional GO.

Unit economics · one venture · Birch

Six percent of a clean claim. Two hundred ten thousand practices.

Target customer

Solo MD practice

Pain

10–12% revenue lost to billing errors

Our cut

6% commission on recovered claims

ARR per practice

$28 K

Addressable practices · US

210 K

3-year SOM

$144 M

Y1 ARR target: $360K · break-even month 8 · build envelope $3,750.

Portfolio math · honest cohort

Most trees will fail. The five that work pay for the rest.

If we launch 52 in Y1 (the aspirational bull), here's how it actually plays out by Y3.

Killed by month 6 · no PMF

~60% · 30 trees

Modest · break-even tail

~30% · 16 trees × $300K ARR

Real winners · Y3 ARR each

~10% · 5 trees × $5M ARR

Y3 floor (cohort math)

~$30M ARR

Bull case (failure rate drops to 40% as the library matures): $50M floor with the same 5 winners pulled up to $8M each → $70M+ portfolio ARR by Y3.

Build envelope $1.7–4K per launch · operating $50–500/mo per tree · 2 humans on payroll. The cost of 30 failures is a rounding error against 5 wins.

The one KPI that matters

First paying customer
in the portfolio.
Q3 2026.

Until then, everything in this deck is theoretical. The model is sound. The math is plausible. The agents work. But nothing is proven until someone pays. We're betting we hit it inside one quarter.

How each tree makes money

Three patterns. Mostly recurring.

Subscription

Flat monthly fee for an agent fleet that replaces a function.

$99–$999/mo depending on industry. Example: Yew at $399/dealer/mo. Cypress at $99/mo flat. Predictable, sticky, easy to forecast.

Trees: Yew · Cypress · Maple · Den · Cradle · Aspen-tier verticals

Commission

A cut of money the agent recovers, books, or saves.

4–15% of the relevant transaction. Pay-when-we-win. Example: Birch 6% of clean claims. Aspen 15% of utility savings. Lane 4% of freight booked.

Trees: Birch · Aspen · Lane · Doorway · Sourcer · Settle

Per-transaction

Volume-priced per action the agent fleet completes.

$0.85–$2,500 per unit, depending on stakes. Example: Sentinel $0.85–2.40 per KYC check. Permit $300 per filing.

Trees: Sentinel · Permit · Sorbnode

Why now

The pieces are all here. The studio is the assembly.

Frontier inference is cheap. Sonnet at <$0.01 per turn, Haiku at <$0.001. A customer-support agent runs all day for less than lunch.
Distribution is automated. SEO content + cold-outreach are now agent-native workflows. No more 5-person growth teams.
Infrastructure is one keystroke. Cloudflare Workers + Pages + D1 + Stripe. Spin up a venture in hours.

The verticals are sitting there. 28 industries already in the queue. Solo doctors, indie e-com, freight, recruiting — all paying for software that doesn't work.
Regulatory clarity is improving. Marlow's office knows what's safe to ship and what needs outside counsel. Faster than it was even six months ago.
The library compounds. Each build makes the next one faster. The advantage widens every Monday.

Risks · three to manage · one we accept

What can break this. And what we already do about it.

The accepted risk

~60% of trees will fail.

Not a bug — a feature of the cohort model. The math survives because 5 winners out of 52 at $5M+ ARR each pay for the 30 that died. We model this from day one. If we don't accept it, we won't ship enough trees to get five winners.

Risk 1 · managed

Agent reliability

Agents drift. Tasks fail silently. Mitigation: Auden's 4 monitor agents, manual approval gates on sensitive outbound (Sorb's pattern propagated), tuesday-rehearsal QA suite before every launch.

Risk 2 · managed

Regulatory tail risk

Healthcare, fintech, cannabis, legal — all have real rules. Mitigation: Marlow's office (9 agents + outside counsel on standby) gates every venture with Go / Conditional / No-Go before Atlas builds.

Risk 3 · managed

Single-operator concentration

Pete is the only decision-maker. Mitigation: Ash (Chief of Staff agent) holds the day. Tod + Jordan vote on every venture. Hard cap: ~25 min of Pete time per launched company. We're also planning to back-fill the COO seat in year 2.

The ask · who we want to hear from

Three asks. In order of urgency.

Customers

First five paid pilots

Healthcare billing (Birch). Cannabis ops (next slot). Spot freight. SMB recruiting. SFR property management. If you run one of these, we want a call.

Advisors

Operators with domain depth

Jordan is the model — CEO of his own thing, votes on every venture in our vertical (cannabis ops). We want three more like him across regulated SMB, fintech, healthcare ops.

Capital

Year-2 scale round

Pete is funding the first ten launches from prior exits. Around tree #10, we'll raise a meaningful round to scale Hollis's intake (more scouts, more verticals) and back-fill the COO seat. Not yet — but talk to us.

The studio

Same playbook. Different tree.
One company a week.

— Old Growth · oldgrowth.ai/console