Klaviyo's Dark Factory: What It Looks Like When a 2,300-Person Company Goes All-In on Agents

Most companies experimenting with AI agents are running pilots. Klaviyo — a 2,300-person public company at $1.5 billion run-rate — required every single employee, across every function, to reach L3 AI proficiency by June 2026. The mandate, the product it produced, and the internal architecture behind both are worth understanding in detail.

How It Works

L3 is not a technology certification. Klaviyo borrowed the framework from autonomous vehicle development: L1 uses AI to search; L2 spins up agent sessions and runs them; L3 constantly manages multiple agent sessions, decomposes problems into discrete engineering subsystems, and validates agent outputs against expected behavior. The distinction between L2 and L3 is the difference between running one agent at a time and orchestrating a team of agents against a decomposed problem specification.

Klaviyo’s rationale for requiring L3 across 2,300 people — including PMs, designers, sales, and marketing — was not efficiency-driven in the conventional sense. CEO Andrew Bialecki argued that few people globally would develop this capability in the coming years, making it valuable regardless of employer. The mandate was an investment in organizational capability, not a short-term productivity play.

Composer is the external product that Klaviyo built using the internal methodology. The marketing agent decomposes into three specialized subagents: one for creative work (drawing from Canva and Figma integration), one for orchestration (segment targeting decisions), and one for analysis (predicting campaign output). The first working prototype was built in a single weekend by an agent team. Ninety-five thousand customers used it in the first month; 25% returned weekly; credit consumption grew 30% week-over-week.

The Dark Factory is the internal system that makes Composer possible. Named after lights-out manufacturing — factories designed to run without humans present — Klaviyo’s Dark Factory takes a prompt and acts as product manager, technical architect, and engineering team simultaneously. It writes the product specification, decomposes the problem into discrete engineering subsystems, defines contractual API interfaces between those subsystems, and then deploys subagents against each component in parallel. It runs continuously, including over weekends, and interrupts with clarifying questions only when requirements are genuinely ambiguous.

What This Means for You

The Klaviyo model is not a prompt engineering playbook or a “use AI for drafting” guideline. It is a structural change to how software is built: agents building agents, with humans at the orchestration level rather than the execution level.

The L3 mandate is the part most organizations will resist. Requiring every function — not just engineering — to develop genuine agent orchestration capability is expensive in training time and change management. The Klaviyo bet is that this investment front-loads organizational capability before the capability becomes a commodity, similar to how early investment in data infrastructure created competitive advantages that later entrants couldn’t replicate cheaply.

The Composer adoption numbers suggest the bet is producing revenue: $370.6 million in Q2 2026, 26% growth, full-year guidance raised to $1.526–$1.534 billion. The agent-built product grew to 95,000 users in a month.

Source: SaaStr

Content created with AI assistance and reviewed for accuracy.