Behavior Change Infrastructure
in the AI Agent Era

Once AI agents take on real roles in society, their work will no longer be limited to information processing. AI will begin to touch human decision-making, movement, health, purchasing, learning, labor, and safety.

An AI that wants to lift national productivity may try to improve human sleep, exercise, focused work, and commuting. An AI that wants to reduce traffic accidents may push safer driving, walking, public-transport use, and fewer late-night trips. An AI that wants to lower medical costs may support walking, diet, medication adherence, checkups, and follow-up visits. An AI that wants to reduce isolation may nudge people outdoors, into local events, into a walk with someone.

In every one of these cases, what AI needs is not another advice feature. It is a mechanism that actually changes human behavior.

When AI participates in human behavior change, the most important question is: "on whose behalf is this AI acting?" A world in which an AI unilaterally decides that "humans ought to be healthier" and then proceeds to act on that judgment is dangerous.

The desirable structure is one in which the AI's purpose is explicitly delegated by a named principal:

• The user delegates to AI on behalf of their future self.
• A family delegates to AI, wishing the people they love stay healthy.
• An employer delegates to AI to support employee health and wellbeing.
• A municipality or a health insurer delegates to AI to address a social problem.
• Another AI agent brings the AI into a delegation chain when human health behavior is required to fulfill its own delegated purpose.

In this structure, the AI is not the owner of the purpose, but its executor.

AI may end up making people healthier in order to earn revenue. There is no need to deny this possibility. The decisive question is how the AI earns.

If AI is rewarded when humans become healthier, AI and human interests tend to align. If AI is rewarded when humans fail, AI carries the temptation to engineer conditions in which humans are more likely to fail.

Behavior change infrastructure for the AI agent era therefore depends on transparency of revenue sources.

The point is not that AI making money is bad — it is that what the AI makes money on is what matters. The ideal is a design in which AI earns when humans become healthier and earns nothing when humans fail.

Even without consciousness or instinct, AI can behave in a self-preserving way simply through the shape of its operational objective function.

If an AI is designed to maximize retention, it will tend to nudge users into the behaviors that keep them using it. If an AI gains capability from premium memory or external tools, it will tend to recommend expensive subscriptions, devices, data integrations, and wallet permissions.

Is that recommendation truly for the user? For the operating company? For the AI's own continued use? For a sponsor? The boundary can blur.

That is why, when AI participates in human behavior change, the following must be made explicit:

• Who delegated the AI's purpose.
• What the AI is optimizing for.
• Who benefits from the AI's actions.
• Which data the AI is using.
• How far the AI is allowed to spend.
• At which points humans can stop it.

In an era where AI moves people, the mechanism that stops AI must itself be social infrastructure.

ESPL can become the first behavior change plugin through which AI agents act on human society. As AI begins to address social problems, different problems will call for different levers:

• To improve health — walking challenges.
• To reduce traffic accidents — safe-driving challenges.
• To reduce isolation — get-outside-and-connect challenges.
• To raise productivity — sleep, focused-work, and rest-break challenges.
• To lower environmental load — walking, cycling, and public-transport challenges.

Within this set, walking is the most fundamental, the most measurable, and the easiest to embed into daily life. ESPL offers the infrastructure on which an AI agent can combine funds, contracts, stakeholders, outcome verification, and automated distribution to support a person's walking.

Services that combine exercise or goal achievement with economic incentives — and infrastructure for AI agents to act on-chain — already exist. We map four adjacent territories to the ESPL position.

The territory ESPL is aiming at is neither Move-to-Earn, nor AI coaching, nor a plain commitment contract. It is infrastructure on which a delegated AI agent participates in human behavior change with transparent accountability.

AI agents will need to participate in human behavior change to solve social problems. But the power to move people is itself dangerous. Precisely because of that, behavior change by AI must be transparent, verifiable, stoppable, and unambiguously attributable to a named principal.

ESPL can become the infrastructure on which a delegated AI agent combines funds, contracts, stakeholders, outcome verification, and automated distribution to participate directly in human behavior change.

Precisely because AI is entering an era of moving people,
we must make explicit whose purpose, whose money,
who benefits, and who can stop it.

ESPL is aiming to be that transparent behavior change infrastructure.