Infinite Garden

Research date: November 2026 Website: infinitegarden.com (earlier domain: infinitegarden.ai)

TL;DR

Infinite Garden isn’t an “agentic commerce” company in the way the term is usually used. They aren’t building rails for AI agents that shop, pay or check out. They’re building an AI system that runs the supply side of consumer goods: it creates and manages a portfolio of brands. The autonomy points at what gets made and sold, not at the buyer.

Company snapshot

Evolution: They appear to have started as a creator-brand supplement studio and moved toward an AI-run brand portfolio.

Carrington’s own framing: He describes the company as architecting an autonomous system to solve performance marketing for thousands of brands at once (“AI-powered, agentic growth marketing”). An earlier hiring post describes building a social-commerce growth engine with creative testing across dozens of creator partners.

Not found: other founders, named brands, disclosed metrics, round size.

The thesis

Both essays were published October 20, 2026.

1. “Demand is a complex system”

Core claim: Consumer demand is complex and irreducible in the complexity-theory sense, not merely complicated. Most methods for finding product-market fit were built for reducible problems, which is why fit is hard to find and hard to keep.

Four reasons research and A/B testing fall short:

1. The problem can’t be fully defined before acting on it. Intervening is what generates the information. Their example is New Coke (1985): taste tests with about 200,000 people favored a sweeter formula, but couldn’t measure people’s attachment to the original. The original returned 79 days later.
2. There’s no clean trial and error. Markets remember earlier offers, and competitors react to earlier moves.
3. Success changes the conditions that produced it. Customers learn and competitors copy.
4. The company is part of the system it studies. They consider this the most important of the four.

Response, “product coevolution”: No A/B-test regime. Instead, constant adjustments that move a whole portfolio forward across a changing terrain. Decisions are better or worse for a given moment, not right or wrong forever. The goal shifts from optimizing to satisficing.

2. “Commercial intelligence with infinite memory”

Core claim: Most companies learn badly in one of two ways:

• Forgetting: findings live in people and slide decks and leave when they do.
• Turning one result into a rule: one winning ad becomes “people want this message,” and the company ends up confidently wrong.

Their method: “evidence, not verdicts” (effectively Bayesian):

• Form beliefs from everything already known before a new result arrives.
• Update each belief only as far as the evidence supports.
• Check beliefs by backtesting: predict results not yet seen, then compare.

“Infinite memory”:

• Nothing is discarded, including null results and lessons from pruned brands.
• Every belief is stored with its conditions: who it applies to, in what situation, on what evidence.
• A belief travels only as far as its evidence. A finding about stressed students isn’t assumed to hold for people in their sixties worried about brain fog until a small test confirms it.
• Minoxidil → Rogaine example: a side effect that didn’t matter for blood pressure became the product. Observations that look irrelevant today can settle a question years later.

Three requirements they name:

1. Evidence from far beyond their own results: trends, market signals, performance data, scientific research, visual styles. Their own data gets no extra weight.
2. A calibrated sense of how much each result should count.
3. A record of why each decision was made, including for pruned brands.

How it works in practice

• Portfolio over single bets: many small moves, applying an investor’s logic inside one company. Some brands get room to grow; others are pruned.
• One integrated loop (“commercial intelligence”): research, product development and brand design run as one process, steered by market response. All brands are linked so what one learns reaches the rest.
• “Make atoms behave like bits”: inventory is the costly commitment, so the largest physical commitments are delayed as long as the evidence allows.
• Long-tail economics: a small brand can be a good business at modest growth if it’s cheap to run. The aim isn’t unicorn brands.
• Decision provenance: reasons are kept alongside outcomes.

Assessment

What it is: In effect, a holding-company or brand-studio model (in the style of Thrasio or Pattern) rebuilt around an AI decision engine, presented in scientific language (complexity theory, Bayesian updating, backtesting).

Potential moat: Data that compounds across brands. Each launch should be cheaper and better informed than the last, which no single DTC brand can do.

Risks and open questions:

• The essays disclose no metrics, brands or results.
• Supplements are heavily regulated and claims-sensitive, which makes them a hard category to automate.
• Portfolio learning depends on attribution quality, and paid-social attribution is notoriously noisy.
• The team’s visible strength is AG1-style performance marketing. The “research lab” framing still has to prove itself on the model side.

Parallel worth noting: “Evidence with conditions attached, plus a record of why each decision was made” is the same problem execution agents face in DeFi: keeping strategy outcomes tied to the regime that produced them, instead of turning one good backtest into a rule.

Sources

• Infinite Garden – home
• Commercial intelligence with infinite memory
• Demand is a complex system
• PitchBook – Infinite Garden profile
• RocketReach – Infinite Garden
• Village – Infinite Garden
• Matt Carrington – LinkedIn
• McIntire – Matt Carrington, CMO of Athletic Greens