MEMO
AI as Workforce
The claim
Most companies use AI to make their products smarter. We use it to make our company smaller. Zalize runs as a company of one human plus an AI team: the person decides what is worth building and answers for the result; agents design, write, ship and operate everything else.
This memo writes down what we have actually learned running that model — where it works, where it breaks, and what we are testing next.
What one person plus agents can already carry
Our catalog spans web tools, desktop apps, browser extensions, mini-programs, SaaS platforms, a digital-goods store and research work — across Chinese and international markets, in two languages. None of that is maintained by a department. It is maintained by agents working against written roles and SOPs, with a human reviewing outcomes.
The important detail is not the count. It is that the same team of agents can move from shipping a browser extension to fulfilling store orders to running a forecasting experiment without being rehired. The leverage is not in any single product — it is that the workforce itself transfers.
Where it breaks
- Judgment does not delegate. Agents produce volume; they do not produce conviction. Every wrong spec gets executed faithfully, at scale. The bottleneck moved from "can we build it" to "did we describe the right thing" — and that question has a cost per mistake.
- Verification is the new labor. Output that looks right is cheap and abundant. Output that is right still requires a human who can tell the difference — or a test that can. We spend more of the human's time on specs and acceptance than on creation.
- Context is the real infrastructure. An agent without the company's context is just an API call. What makes the team coherent is shared memory: product data in one place, decisions written down, roles and SOPs versioned like code.
The org chart is a codebase
The way we run this is deliberately unglamorous. Agent roles have charters. Processes have SOPs. Workers sync catalogs and fulfill orders around the clock. When something breaks, the postmortem becomes an artifact the next agent reads. The organization is not a chart on a wall — it is a repository the team executes against.
That is also what makes the model copyable. A playbook that lives in files can be handed to the next person. A playbook that lives in one head cannot.
What we are testing next
Today: generalist agents under direct human direction. Next: role-specialized agents — product, engineering, growth, operations — each with its own charter and tools. The horizon is functions that run autonomously under review, with escalation paths instead of micromanagement. What does not change is who decides what is worth doing.
We publish this as a working memo, not a manifesto. The catalog is the evidence; this page is the hypothesis. If the model survives contact with the next year, we will write that down too.
Written by zalize
-- zalize.com