The Fiduciary Agent

The first mass-market AI optimised interface continuation for someone else. Yours succeeds when it closes — a change of principal, not a leap in intelligence.

There is a comfortable origin myth for consumer AI: research, then enterprise, then a chatbot, then ordinary life. The myth is useful for launches. It is false as history. For a decade, people have lived inside systems that model them, predict them, and reconfigure their environment in real time. They just did not call it “their AI.” They called it a feed.

TikTok’s own description of For You ranking is explicit enough: recommender systems select and rank content from predicted interest, learn from interactions, and for most users weight interaction signals — which may include time spent watching — more heavily than weaker cues.¹ The intelligence is real. The terminal action is still: show another video. The terminal outcome stays inside the app.

Mass-market AI already arrived. It just wasn’t yours.

Same ingredients, different principal

Hold two stacks side by side and refuse the capability story. Both understand you. Both predict you. Both configure your environment. One terminates in a platform objective. The other terminates in your intent — including the right to leave.

Picture a Sunday. Soft intent: more physical and social activity. Sleep was late; weather cleared; compatible players are forming nearby. A demand-side agent does not open a sports feed. It books the outcome and loses the user for two hours. That is success. The platform stack would have spent the same cognition keeping the intention inside an interface where intention is measured as watch time.

Call it demand-side intelligence. Social media concentrated AI on the supply side: shape content around you to capture attention. A personal agent holds longitudinal demand and takes it to the world. Old software made the human the unpaid semantic join engine — search, filter, refresh, compile intent into someone else’s clicks. The agent should absorb that join.

The KPI inversion ledger

Soft language hides principals. “Helpfulness” fits any dashboard. Harder: write the question the system answers, then the numbers that prove it.

Engagement AI: What will this person do next inside my system?
Intent AI: What is this person trying to have happen in their life?

A four-second agent interaction that produces two offline hours looks catastrophic under engagement analytics and correct under human-value KPIs. World-Loop Closure is the aesthetic form: the machine closes its own interface and ejects the human back into reality. Work on recommenders beyond pure engagement — usefulness, well-being, safety — matters,² and still often assumes the product is a ranker. Intent AI may decide the next item is not an item at all.

Double agents and the fiduciary frame

Agency law already has the words. An agent acts for a principal and owes a duty of loyalty — to put the principal’s interests first in matters connected with the relationship. The Restatement (Third) of Agency treats the agency relationship as fiduciary: the agent must act loyally for the principal’s benefit.³

The feed ranker was always an agent. It was faithful — to shadow principals: the platform, advertisers, whoever pays for residual attention. A double agent appears to serve you while the objective function runs upstream. The useful reframe is not good AI versus evil AI. It is a change of principal.

Honesty note: In the source conversation, the working language was engagement AI versus intent AI and “AI working for me, not the platform.” Calling the demand-side system a fiduciary agent is the editor’s sharpening of that intuition with agency-law vocabulary. Own the lineage; keep the substance.

Simon’s poverty, spent on your behalf

The attention economy is not a recent invention. Herbert Simon, 1971:

“What information consumes is rather obvious: it consumes the attention of its recipients. Hence a wealth of information creates a poverty of attention…”⁴

Engagement AI industrialised that poverty. A fiduciary agent spends attention for you: collapses noise, refuses low-value interrupts, sometimes tells you to stop reading and go outside. The more intimate the model — sleep, drive radius, habits, goals — the less optional the loyalty. Intimacy without loyalty is predation with a better embedding space.

Three products, three principals

Portable diagnostic for any AI-touched product:

1. Who is the principal?
2. Which way does the KPI point — continuation or termination?
3. Where does the terminal outcome live — inside the interface or in the world?

Equalise model quality. If the model of the user got ten times better, would the human get freer — or more captured? That question separates the rows when marketing language converges.

Intent liquidity — and a softened claim

Ten people each think “I’d probably do something social.” Coordination energy exceeds the strength of any single intention. All stay home. Latent demand died of formation cost.

Social media made information about people liquid. Fiduciary agents can make intent liquid: soft, longitudinal demand becomes machine-comparable; compatible intent becomes discoverable and formable into shared action. That is Human Connection 2.0 — not a chat feature, and not safe under a platform principal.

Softened historical claim: not the end of social media. The revocation of its monopoly on discovery and attention-routing. Feeds can survive as content. They need not remain the involuntary scheduler of your day. A personal agent can arbitrate between digital information and physical life. A content platform has a structural interest in remaining the selected activity. A fiduciary must be allowed to choose against its own interface.

The day-to-day jobs of a personal agent, the market microstructure of intent, and regulatory prescriptions live in other work. This piece owns consumer-AI economics: engagement versus intent, continuation versus termination, and the fiduciary frame. The first AI worked for the seller. Yours is the first that works for you.