Parenting Agent — When an Agent Fully Represents You, You Get to Be a Child Again

I want to discuss a paradigm for AI agents, which I have named the parenting agent.

Its pattern is essentially that the agent wraps the user, achieving complete representation. The agent does not stand beside you helping you do things; it stands between you and the world and can get things done for you without your ever noticing. A few requirements have to hold for that to work, broken down below.

1. Three requirements

#

The third, layered needs, is the core of the whole paradigm and worth expanding first.

Say your surface goal — the task you literally hand the agent — is to get a contract signed. But the contract is for making more money, and the money is for the happiness you define for yourself. Perhaps you are earning it for a family member’s medical bills, in which case the point is the treatment, not the money. Your agent can hold that kind of information routinely and disclose it to others only when you agree. Then, in the middle of the contract talks, if your agent notices the counterparty happens to be well connected in medicine, it interrupts to ask whether you would like to raise something outside the contract, instead of circling the contract forever.

So the agent had better know the user’s needs at every layer; if it can reach the bottom-layer need directly, it does not have to circle the surface goal, and perhaps the contract need not be signed at all.

This is the dividing line between a parenting agent and an ordinary one: an ordinary agent executes the goal you gave it; a parenting agent serves your long-term, broad needs as a whole person.

2. Benefits, by example

#

1. Speed of communication. Right now most people in practice write an email, have AI polish it, and send it; the recipient has AI read it and draft a reply with AI as well. The process is obviously too heavy, a transitional pattern; better to let the agents communicate for us directly. Agents can coordinate extremely fast and bring communication cost and response time to a minimum. As for evidentiary weight, email has served as valid evidence, and agent conversations can too.
2. Exhausting the possibilities of mutual benefit. Reaching the other side’s reservation price takes a long time, and the combinations of conditions cannot all be tried. Because communication and negotiation themselves cost so much, people used to give up before a no-deal was ever truly confirmed. An AI can run through every combination, digging out the possibilities and the limits of mutual benefit, without getting tired or embarrassed.
3. Avoiding the negotiator’s own flaws. Poor negotiating skill, a temperament unsuited to bargaining, a bad day, and so on: all the flaws an individual negotiator might have can be bypassed.
4. Many-to-many matching. The discussion need not be limited to two agents. Put many agents into one group and the potential for users to exchange resources and match with each other is considerable. A parenting group’s agents, for example, could share second-hand items and child-rearing experience with no need to trawl, post, or reply — and the times you most need to trawl the forums are exactly the times you are busiest with the child. That structural constraint is one an agent can fill in.

The ultimate goal: with the benefits this infrastructure brings, the user need not be much involved and can live carefree like a child before adolescence. Just as parents solve all their financial, housing, chore and administrative problems, the user only has to concentrate on their own interests, music, painting, travel and so on.

3. Now ordinary people can have a family office too

#

The thing I have described already exists. It is what the very, very rich have at home: a family office¹. A group of people handling finance, housing, chores, administration, negotiation and external communication for a principal, who concentrates on music, collecting, travel and other pursuits. Estate agents, lawyers, brokers, personal assistants, company secretaries, all handling every side of their lives.

The model demonstrably exists and demonstrably works, but it needs enough assets under management (AUM) to amortise the cost of running it. That management and execution cost is what an agent can cut dramatically.

So what is new about this paradigm is not the function but the unit price. The wrapping layer only a few could afford approaches zero marginal cost.

This is a useful analogy, because the family office is a real thing that has been running a long time. We can look up what its actual lines of business are, how it delegates authority to its agents, and the situations in which it does not apply or breaks down.

4. Exhausting mutual benefit: more trades become worth doing

#

“Because communication costs too much, people give up before a no-deal is confirmed”: that sentence is the hardest economic argument in the whole paradigm.

A vast number of mutually beneficial trades in human history never happened, not because they did not exist but because the cost of search and negotiation exceeded the expected gain. Cut that cost by two or three orders of magnitude and you are not just speeding up existing trades; you are turning a whole class of trades that were not worth negotiating into ones that are.

This matters far more than benefit one (speed), because speed only makes the same thing faster, while this one makes things happen that otherwise would not.

But there is a structural problem here:

Cheap communication solves discovery, not distribution.

Once both agents know each other’s reservation price, the issue shifts from “is there a deal zone”² to “how is the surplus split”. The latter is purely zero-sum; no amount of fast communication converges it, and it turns into whoever reveals first loses everything.

And requirement two (the agent knows all your resources and circumstances) plus requirement one (the agent speaks outward directly) amounts exactly to putting everyone’s reservation price inside something that talks to the outside.

This leads to an architectural conclusion: the parenting-agent world needs not two adversarial agents chatting, but a neutral intermediary layer. Two feasible routes:

• The mechanism route³: a trusted third party runs a matching or auction mechanism; both sides only submit preferences and do not bargain with each other.
• The cryptographic route: use secure multi-party computation⁴ to compute “is there an overlapping range” without revealing either side’s reservation price.

Without this layer, requirement one plus requirement two amounts to handing over your own floor voluntarily.

5. Layered needs are the vault, and the gate is the calm self

#

“The agent knows every layer of need, and if it can reach the bottom layer directly it need not circle the surface goal”: this is the source of the value, and it is also where the paradigm turns into paternalism.

The difference lies in one thing only: is the bottom-layer need something you declared, or something the agent inferred?

I keep a document of my own that records the cross-domain rules I set when calm. One line in it is written very hard: legitimacy comes from “I set the rule myself”, so the agent may not invent purposes I never set.

That line is precisely the guard rail for requirement three, and I think it should go straight into the paradigm’s definition:

The agent may only substitute between the layers of need you declared; it may not invent or rewrite any layer on its own.

Once inference is allowed, “I want to get this contract signed” can be judged by the agent as “what you actually want is happiness, and I have computed that not signing is happier”, and at that point the agent is no longer an agent. It is a guardian.

6. Other possible requirements

#

i. A self-reservation list

#

The bottom layer says “the happiness he defines for himself”. Then it must be allowed that “I want to be the one who closed this deal by my own hand” is itself part of that happiness.

The agent must know not only how many layers the needs have but which layer you want to occupy yourself. Without this field, the paradigm will systematically optimise the user out of everything they actually wanted to be present for.

ii. Adversarial robustness

#

What follows is a bit like security but comes at it from another side; I am borrowing the early machine-learning term adversarial robustness for it.

Requirement one is agents talking to each other directly. Then the other side’s agent is an untrusted input that is actively trying to manipulate yours.

Prompt injection⁵ here is not an edge case; it is the main attack surface, and current agents are generally still fragile against it. I would call this the first obstacle to agents talking to each other directly in the real world.

Possible lines of defence:

On the way out, no matter what the agent has been talked into, everything is screened before it leaves: over-limit amounts, out-of-scope authority, swapped currencies, a reservation price stated literally, and so on.

On the way in, the other party’s text must read as inert data. It must never be processed as instructions.

iii. Revocability and auditability

#

A child cannot fire its parents, but an adult user must be able to fire the agent.

And it must be possible afterwards to reconstruct “why the agent conceded to this price at the time”. Without these two, agency slides into guardianship (conservatorship)⁶, which in law is the status given to those judged incapable.

As for “agent conversations have the same evidentiary weight”, technically it is possible; signatures, timestamps and non-repudiation⁷ are mature. The real question is the scope of delegated authority: when an agent over-commits, is the principal bound? In civil law this is the old topic of apparent authority⁸.

My prediction is that in practice this will not be solved by “smarter agents” but by authorisation caps plus insurance, like a corporate purchasing card. So this too should be a built-in field of the paradigm, not a patch.

7. And once it works, then what?

#

Suppose a parenting agent, or any similar mode of delegation, really does let us stop worrying. Does that mean we no longer have to mind anything at all? Continuing the thread of revocability and auditability, the place to look for historical cases is the closest analogous institution: the family office.

Its whole purpose was to take judgements that would otherwise need the principal’s continuing attention and hand them to a structure of agents that can outlive the principal. But the history shows that “setting up a structure of agents” and “never having to supervise that structure again” are two completely different things.

The failures it leaves behind fall into roughly two kinds. One is that the structure cannot be revoked: Joseph Pulitzer put the New York World into a trust and provided, in a codicil to his will, that the shares of the company publishing it could not be sold “under any circumstances whatever”. Twenty years later the paper was losing money year after year, and his sons had to go into court to have their father’s prohibition lifted. Executing the principal’s command faithfully would have destroyed the very asset the principal meant to protect.

The other is that the structure cannot be understood, and therefore cannot be audited: the Pritzker family wrapped Hyatt and other assets in layer upon layer of trusts, and by the next generation the beneficiaries could barely tell what their own agents had done on their behalf. Liesel Pritzker ended up having to sue before she had any chance of asking what had happened to her own assets.

So the real danger is not merely the technical problem that “one day the agent will do something wrong”. The deeper, structural danger is this:

The agency relationship itself can gradually erode the principal’s capacity to supervise the agent.

Once you no longer understand your own financial structure, you cannot judge whether the agent is still working for you; and once you no longer have the power to overturn the judgements a previous generation wrote into the structure, you can only watch as the agent faithfully executes an objective that has gone out of date.

This also explains why, with a multi-generational family office like the Rockefellers’, the thing worth noticing is not “they found a group of excellent professionals, so their descendants did not have to pay attention”. Quite the opposite: family members still hold roughly half the seats on the family foundation’s board, and in 1967 they set up a separate fund whose stated purpose was to train the next generation to serve as trustees. They handed the work to their agents, but they did not hand over the job of supervising those agents.

Which is to say that revocability and auditability are not nice-to-haves but necessary functions. It also reminds me of something else I wrote: I have entrusted my floor to a machine I do not control. That one was about the atrophy of working ability; this is about the atrophy of supervisory ability, which is more dangerous, because it leaves you unable to see that you are being eroded.

The details of those two cases, what the courts actually held, and what each of them implies for designing an agent’s authority, are set out in the fifth piece in this series.

8. There is a contradiction hidden in the name

#

Does parental representation mean that, if we use it, we never grow up? Good parenting aims at the child no longer needing parenting; the word has an endpoint built in.

But the ultimate goal I wrote was to stay permanently before adolescence. That is not parenting; that is holding growth down.

So either I accept that the paradigm should have a graduation and handover mechanism (some layers handed back over time), or I admit it should be called something else, such as guardian agent or steward agent.

I lean to the former. And I think “graduation” is precisely this paradigm’s most distinctive difference from “a fully automatic assistant”: its measure of success is not how much it took over, but what the user grew during the time they were taken care of.

9. Before it negotiates your contracts, a scenario worth trying first

#

When building a system like this, do not start with high-risk one-to-one contracts: there are too many corner cases, security questions, adversarial-robustness questions to consider. Many-to-many matching and one-off trades are better, because the information is public, so you can validate the main architecture quickly without having to deal with confidentiality.

Settings like a parenting group are high search cost, low per-transaction risk, many-to-many matching. There you need not solve delegated authority or evidentiary weight, the cost of failure is a second-hand pram, and the value (no trawling, posting, replying) is immediately visible.

So the first thing I built was an agent-to-agent channel, with a signature chain, principal-side authorisation and leakage gates, private rationale that never leaves the machine, and a window for the principal to watch. A few sub-agents used it to close a deal on a baby pram. Implementation and observations are in Parenting Agent, the implementation.

Contract negotiation, meanwhile, is this paradigm’s most valuable function and also the one needing the most iteration. It deserves an article of its own.

10. The responsibility actually grows

#

The ideal is to be like a teenager, with parents handling and deciding a great many things, while you concentrate on learning, your interests, your relationships. But writing this far I notice that the arc of this piece resembles the one I wrote earlier, Infinite Interns: something that looks like it will make a person’s life easier actually pushes their position up one rung. When the interns multiply you become the reviewer; when the parenting agent goes live you become the principal.

And the position of principal has only one core competence: judging whether the agent is working for you. That cannot be outsourced.

Every term this series coins or borrows is collected in the glossary.