Who Manages the Agents?
2026-07-12 · 5 min read · Janaina Maia
This week, three stories sat next to each other and said something none of them said alone.
OpenAI posted a job for a product manager dedicated to families, caregivers, and older adults. ChatGPT's audience is aging up — users 35 and older grew from 26% to 31% in a year, while the 18-24 bracket dropped from 34% to 29%. Nearly one in four American parents used ChatGPT last quarter. OpenAI is designing for households, not just individuals. The job posting calls for experience building "trust-sensitive consumer experiences."
Meta launched a feature letting anyone @-mention a public Instagram account to generate AI images using that person's photos. No notification to the person. No consent mechanism. Talent agencies and users revolted. Four days later, Meta killed it. "We've heard the feedback that this feature missed the mark," they said.
And Gavriel Cohen, a startup founder, published a piece called "Don't Go Quietly Into the AI Night" that argues the real AI divide is not who has access, but who manages the agents. His thesis: the frontier is racing ahead but the median person hasn't moved. Less than one percent of developers are at 100x productivity. Everyone else is still at roughly 1x. The answer, he says, is not better models for the elite. It's giving every person their own agent to manage — and teaching them to lead.
Three stories. One question: when AI enters your life — your family, your team, your company — who manages the agents?
OpenAI's family bet is a design inflection
Until now, ChatGPT has been a single-user product. You open it, you type, you get an answer. It's you and the machine, one on one.
A family product is fundamentally different. Families share accounts, have different ages and permissions, and have trust boundaries that don't exist in a single-user context. A product manager for families isn't adding a feature. They're designing a new category: the AI that lives in your household.
The Family Online Safety Institute just published research showing that parents underestimate how much their kids use generative AI. 27% of parents said their child used AI in the past week. 38% of kids said they did. There's an 11-point gap between what parents think is happening and what's actually happening. That gap is a design problem. When OpenAI adds parental controls, trusted contacts, and family-specific safety routing, they're building the infrastructure for a product that mediates between generations, not just individuals.
But here's the tension: OpenAI is also facing multiple lawsuits from parents who say ChatGPT contributed to their children's harm, including suicide. The same company that wants to be in your household is being told by courts that its product isn't safe for the people already in it. You can't design for families without designing for vulnerability. And vulnerability isn't a niche use case — it's the default state of most human beings most of the time.
Meta's four-day feature is a masterclass in what not to do
Meta's Instagram AI feature was designed backwards. They started with the capability — "we can generate images from public accounts" — and never asked the question that matters: "what happens to the person whose photos we're using?"
The feature let you @-mention any public Instagram account and use their photos as reference material for AI-generated images. The person referenced would not be notified. They could not opt out in advance. They could only disable it after discovering it existed, and even that required knowing the feature was there.
This is the same mistake every platform keeps making: they design for the person using the feature, not the person affected by it. Instagram has 2 billion users. The number of people who could be referenced without consent is orders of magnitude larger than the number of people who would do the referencing. The feature wasn't just tone-deaf — it was architecturally indifferent to consent.
And then there's the detection problem. Reuters reported this week that Meta's own AI image detector fails to identify 55% of its own AI-generated images once they're cropped. The company launched Muse Image with a Content Seal watermarking system that it said could identify its own AI images even after cropping. It can't — not reliably. So Meta built a tool that lets people create fake images using real people's photos, and the tool that's supposed to detect those fakes doesn't work on modified images. That's not a feature. That's a liability with a press release.
What is agent management, really?
Cohen's essay introduces a framework that I think is genuinely useful: every person becomes an agent manager. Not in the sense of becoming a programmer or an AI researcher. In the sense that you already know how to do the things an agent manager needs to do.
You've managed people. You've managed projects. You know how to communicate what you want clearly, how to give feedback, how to review others' work, and how to take responsibility for the output. Those are management skills. And those are the skills needed to effectively manage an AI agent.
The key insight: the median developer hasn't gained meaningful productivity from AI tools. A tiny fraction of power users are at 100x. Everyone else is still at roughly 1x. Making a smaller group more productive doesn't move the average. Making the median person 2x does.
This reframes the entire AI product conversation. Instead of asking "how do we make AI smarter?" it asks "how do we make the median person more capable with the AI we already have?" That's a product design question, not a research question. And it has specific implications:
- Design for the manager, not the replacement. The interface should help the human direct, review, and override — not optimize for the AI doing everything autonomously.
- Give people their own agents. Not agents thrown into shared channels with unclear ownership. Agents that belong to a person, that they manage, that help them do their work.
- Make growth the strategy, not cost reduction. A company whose only AI strategy is headcount reduction is using an exponential technology for a linear goal. The path to 2x your company is making the median person 2x.
- Agents need to be sovereign. Your company must own the agent's identity, memory, skills, and audit trails. If your vendor manages the agents, your company becomes redundant. You've outsourced the intelligence layer.
Why this matters for product design
The three stories this week are about the same thing: the design of power in AI products.
OpenAI is deciding whether ChatGPT serves individuals or households. The answer changes everything — from safety routing to consent to shared memory to parental oversight. If the household is the unit, the product has to be designed for the least powerful person in it. That's a child, an elderly person, someone in distress. Not the power user.
Meta decided that the person using a feature matters more than the person affected by it. They were wrong. The product lesson is simple: design for the person who didn't ask for the feature, because they're the one who will suffer from it.
Cohen's framework is the design brief for the next generation of AI products: put the human at the center. They set the objective. They review the work. They own the result. The agent amplifies. The human decides.
This is not anti-AI. It's pro-human-agency. And it's the design challenge of the decade.
My take
Every AI product team right now is making a choice they might not realize they're making. The choice is between two product metaphors: the assistant and the manager.
The assistant metaphor says AI is there to help you. It waits for your command. It does what you ask. It's polite, competent, and invisible when you don't need it. This metaphor is seductive because it's familiar. Everyone has had an assistant, or wanted one.
The manager metaphor says AI is something you direct, review, and take responsibility for. You set the goals. You check the output. You own the result. The AI does not replace your judgment. It amplifies it. This metaphor is less seductive and more accurate.
OpenAI's family PM is an assistant bet with manager consequences. Meta's Instagram feature was an assistant designed without a manager in sight. Cohen's essay is a manager's manifesto.
The products that will matter are the ones that treat every user as the manager of their own intelligence — not the recipient of someone else's. Not the congregation of a technical clergy. Not the passive beneficiary of an all-knowing system. The person who decides, directs, reviews, and owns.
Who manages the agents? You do. Or at least, you should. The design question is whether the product agrees.