Running seven products with an AI Director

One person, seven live products, and an operating model where AI agents do the work and a written charter decides what still needs a human.

By Parham Modirniya · · 6 min read

I run seven live products: apps on the App Store and Google Play, multiplayer games, a VPN, and an open-source game. I am one person. For most of the past year the bottleneck was not writing code, because AI agents write most of it now. The bottleneck was everything around the code: knowing what each product needed next, keeping seven sets of decisions consistent, and making sure nothing important happened without me knowing.

So I stopped treating AI as a coding tool and started treating it as an organization. This is how that works.

The problem: more products than attention

A single agent session is excellent at a well-described task. It is poor at remembering what was decided last week in a different product, noticing that a change in one place should be applied in five others, or knowing which of today's ten possible tasks matters most. With seven products, I was the only place all of that lived, and I was the limit.

The model: a Director, briefs and reports

The core of the setup is an AI “Director”: an agent whose job is the whole portfolio, not any single codebase. It never ships product code itself. Instead it works through three kinds of written documents:

  • Briefs. The Director writes a brief for a unit of work in one product: the goal, the constraints, what done looks like, and what must come back to me.
  • Reports. A separate agent session does the work in that product and files a report: what it changed, what it verified, what it could not verify, and what it is waiting on.
  • Decision records. When I decide something, it is written down where every future session will find it, so the same question is not asked twice and the same decision is applied everywhere.

The Director reads the reports, answers them, and plans the next brief. Because everything is written, the system has a memory that no single session has, and I can read the state of any product in a few minutes.

The rules: a written standard and charter

Agents are fast and confident, which is exactly why they need rules. A written standard and charter define what an agent may do on its own and what must come to me. In short:

  • Agents may change code, write tests and documentation, open pull requests, and report.
  • Anything that reaches users, money, policy or the public needs my explicit yes. That includes production deploys, legal text, store submissions and pricing.
  • Agents get the least access the work needs.
  • A report must separate what was verified from what was assumed. “Should work” is not a result.

The charter is versioned, and every product carries a copy, so a session started in any product begins with the same rules.

Portfolio-wide work

The model pays off most on work that cuts across every product. One brief can run in all seven at once, with each product reporting back on the same checklist. I have used it for a security and privacy hardening pass across the whole portfolio, for a review that checked each product's privacy policy, terms and store forms against what the product actually does, and for adopting a single standard for privacy-respecting measurement. Done by hand, each of those would have been weeks of context switching.

What works

  • Consistency. A decision made once is applied everywhere, because it is written where every session looks.
  • Parallelism. Several products move forward at once, and I review results instead of producing them.
  • Honest status. Reports that must separate verified from assumed turn out to be more useful than any dashboard.

What doesn't

  • Documents drift. Instructions that were true last month quietly become wrong. Stale documents have to be found and marked, or agents will follow them confidently.
  • Verification is the real cost. Agents produce work faster than anyone can check it. The system is only as good as its tests and its reports.
  • Agents are confidently wrong sometimes. The charter exists because a convincing report is not the same thing as a correct one.

What a human still owns

Priorities, taste, and accountability. I decide what matters this week, what each product should feel like, and what goes out under my name. The Director makes it possible for one person to hold seven products in their head; it does not make the decisions for them.

AI agents can do most of the work. Someone still has to own the outcome, and the operating model should make that easy, not optional.