What if AI improved the systems around us?

Most AI products ask a person to do something: open an app, type a request, follow advice. I’m interested in a different question. What if AI could improve the conditions around people, so their lives got better without requiring them to use it?

That is the idea behind Manager. The direction can be described as world engineering: modeling a shared environment, comparing changes to it, acting through the people or organizations with authority to make those changes, and learning from measured results.

The environment is not just a physical place. It includes what is available, what it costs, who can reach it, when it can be used, and the rules that distribute resources and opportunities. A better service network can change what people can access. A different allocation rule can change who carries a cost. These changes affect many people through a common structure.

This is why the distinction between a personal assistant and Manager matters. An assistant can help someone navigate the world. Manager’s central question is which parts of that world should change. Anicca’s mission is End suffering. Physical health, mental wellbeing, and financial security are outcomes to examine in pursuing that mission.

There is useful work to learn from. Simile and Aaru study how people respond to conditions and decisions. Generative Agents provides an architecture for memory, reflection, and planning. Concordia offers agents interacting through a game master. AI Economist separates economic agents from a planner that changes institutions. AgentSociety connects population, movement, social interactions, and economic activity. These are different pieces of the problem, rather than interchangeable implementations of Manager.

The hard part is evidence. A convincing simulation can still be wrong. A system can predict survey answers without predicting behavior under a new institution. An agent can execute a change successfully without improving anyone’s life.

So I am designing four evaluations separately. Can the system predict responses? Can it choose a better structural intervention? Can it apply the change and verify it? Do independently measured outcomes improve against a credible comparison?

The distinction protects the central claim. Access to a service is not itself a health outcome. A company’s profit is not the same as beneficiaries becoming wealthier. An average improvement can conceal harm to a group. Each claim needs its own evidence, scope, costs, and uncertainty.

A system that acts also needs behavior and results that can be examined. The proposal should therefore be distinguished from demonstrated real-world impact.

The first technical milestone is a reproducible environment with explicit resources and rules, a behavior-model boundary, baseline comparisons, and held-out tests. It can run in the cloud; requiring a personal local machine is not part of the vision. Later, the same action and evidence contracts can connect to an authorized real environment.

No beneficiary login should be needed. But no interface must not mean no accountability. People should be able to inspect what changed, what was measured, what failed, and who benefited. The ambition is broad. The proof has to be specific.

Sources: Generative Agents, individual simulations, Concordia, AI Economist, AgentSociety, Simile, Aaru. References do not imply affiliation or endorsement.