a builder's codex
codex · operators · Amol Avasare · ins_cash-four-stage-growth-automation

Automate the four stages of a growth experiment; keep humans on alignment

By Amol Avasare · Head of Growth, Anthropic (ex-Mercury, MasterClass) · 2026-04-05 · podcast · Anthropic growth, CASH, and the squeezed PM

Tier A · TL;DR
Automate the four stages of a growth experiment; keep humans on alignment

Claim

Run growth experimentation through a four-stage substrate — identify opportunity, build the change, test against a quality + brand bar, ship and analyze — driven by Claude. The fifth stage (cross-functional alignment) stays human, and that is the lasting bottleneck.

Mechanism

Most growth experimentation is loosely coupled steps that already have rich playbooks: ideation, implementation, QA, analysis. A capable model can drive each step end-to-end against a written brand and quality bar, with current win rates around "junior PM 2–3 years in." The expensive human input is no longer building the experiment — it is the political and aesthetic work of getting six people in a room to agree on what to ship.

Conditions

Holds when:

Fails when:

Evidence

"Identify opportunities → build the feature → test against quality + brand bar → ship + analyze."

"We will have AGI and it will still be impossible to get six people in a room to align."

The team is led by Alexey Komissarouk inside Anthropic. Win rate today is named at "junior PM 2–3 years in." The substrate wasn't viable before Opus 4.5; it is now. Human-in-loop review need is decreasing weekly.

— Amol Avasare on Lenny's Podcast, 2026-04-05

Signals

Counter-evidence

Operators outside frontier labs may not have the brand-bar maturity, the model access, or the org buy-in to run this. Without the codified guardrails, automating stages 1–4 produces fast slop. The win is not in the automation; it is in the prerequisite of having explicit quality definitions.

Cross-references

Open the interactive view → View original source → Markdown source →