The Hypothesis

Can an agent analyse a product's data, identify the highest-impact growth loop, and then build and operate it autonomously?

The Concept

Growth loops are the compounding engine behind every great product — but designing them requires deep expertise and months of iteration. This experiment asks: can an AI agent analyse a product's data, identify the strongest growth loop, and then build and operate it autonomously?

Growth loop machine

The hypothesis

Can an agent analyse a product’s data, identify the highest-impact growth loop, and then build and operate it autonomously?

The concept

Growth loops are the compounding engine behind every great product — but designing them requires deep expertise and months of iteration. This experiment asks: can an AI agent analyse a product’s data, identify the strongest growth loop, and then build and operate it autonomously?

How it works

1. Product data and metrics
2. Analyse current state — user journeys, conversion funnels, retention curves
3. Identify loop type — viral, content, paid, product-led
4. Design loop architecture — trigger → action → reward → invest
5. Build and instrument — landing pages, emails, referrals
6. Operate and optimise — measure, adjust, compound

The agent handles the full cycle — from analysis to execution — then continuously optimises the loop.

What it explores

• Can an agent correctly identify which growth loop type suits a given product?
• How much historical data does the agent need to design a viable loop?
• Can the agent build the actual assets (pages, emails, referral mechanics)?
• Does an AI-operated growth loop compound at the same rate as a human-managed one?
• What happens when the loop hits diminishing returns — can the agent pivot?

What we found

• Loop identification works well — the agent reliably picks the right loop type given enough conversion data
• Asset creation is the bottleneck — the agent can design the loop but struggles with high-quality execution of individual assets
• The agent is better at optimising existing loops than creating new ones from scratch
• The agent tends to pivot too early when results are slow

Learnings

• Strategy and execution require different agent profiles — a single agent that does both will excel at one and struggle at the other.

• Loop compounding needs a patience threshold built into the agent — without a minimum dwell time, early pivots kill loops before they have time to compound.

• Pair a strategy agent with specialist execution agents — the growth loop agent handles loop design while landing page, email, and ad agents handle asset creation.

Where it goes next

The identification and optimisation layers are strong enough to build on. The asset creation gap is where the rest of the Flywheel stack comes in — landing page agents, email agents, and ad creative agents handle execution while the growth loop agent handles strategy.