Living worlds for agents.

What do we mean by a living world?

In this guide, a living world is an environment that changes over time as people, agents and background processes interact. It includes persistent state, rules for change and a defined view of that state for each participant. A place can be represented spatially, through software interfaces, or through both.

The term is a design description, not a universal technical standard. For any particular environment, explain what persists, who acts, how time advances and which interactions are modeled. That definition is more useful to a researcher than the visual complexity of the scene.

Separate appearance from behavior

A realistic image helps a person understand a setting, but behavioral realism comes from the rules behind it. A calendar should preserve a confirmed meeting. A conversation should affect what its participants know. An unavailable resource should remain unavailable until an event changes that state.

The research paper Generative Agents explores one approach: agents in a small simulated town use memories, reflection and planning to produce social behavior. It offers an architectural reference for interactive simulations. Believable behavior in a simulation does not establish that the simulation predicts a real population.

For an evaluation environment, define the specific behavior you need to test before increasing visual or social detail. If the task is rescheduling a delivery, the meaningful details may be opening hours, a delayed notification and a shared address record.

A concrete example: one changed appointment

Imagine a modeled neighborhood clinic. A person changes an appointment on a phone. A scheduling agent sees the update in a desktop calendar. Another participant is arranging transport using an earlier time. The world must determine which records changed, who received a notification and when each participant could have known about the change.

This hypothetical scenario needs several distinct layers. The appointment is authoritative state. The phone and calendar are interfaces to that state. Notifications are events. A participant’s memory is a partial account of events they observed. Keeping those layers distinct makes a missed update explainable.

A useful variation delays one notification while holding everything else constant. Another changes the participant’s availability. These variations test different abilities: noticing new evidence, reconciling disagreement and updating a plan. None requires pretending that a generated person is a measured model of a real individual.

Build the smallest coherent world

State

Store the facts the task depends on: records, resources, relationships and permissions.

Events

Represent changes with their actor, time, affected state and cause.

Observations

Specify what each participant can see through each interface.

Time

Choose whether the world advances continuously, in steps or after explicit actions.

Reset

Restore the initial conditions needed for a fair comparison.

Start with one complete workflow. Add background activity when it introduces a meaningful challenge, such as contention for a shared resource. An event history helps explain a result, while a versioned initial state makes the test easier to repeat.

Evaluate consistency and usefulness

Check whether the environment preserves its own rules. Can two agents reserve the same exclusive resource? Does a completed task stay completed? Can an agent observe information it should not have? These checks test the environment before it is used to judge an agent.

Then evaluate outcomes across multiple scenarios. Retain failures caused by missing information separately from failures caused by invalid actions or infrastructure errors. Report which social behaviors are scripted, learned or human supplied. The interpretation of a result depends on those modeling choices.

Coasty’s direction

Coasty’s platform connects modeled environments, human interactions, computer use and phone use for multi-agent systems. Its cinematic website illustrates that direction; the visual journey is not an interactive autonomous-world simulator. Explore the platform to discuss an environment around the work you need agents to perform.