Real users have higher entropy than simulated ones.

Model-generated users converge on tidy behavior.

Ask a language model to simulate many users and the outputs may look diverse at first. Names, phrasing, and superficial preferences change. But the underlying behavior often collapses into a few polite, coherent modes: read the instructions, follow the expected flow, and stop when the stated goal is complete.

Human behavior has a much wider distribution. A person may paste an entire document into a tiny field, drag an object outside its intended area, click twice during a transition, or return days later with a different objective. These actions are not rare noise around a canonical path. They are part of the environment a reliable system must survive.

Realism is not visual fidelity. It is behavioral coverage.

Opening the application is not using it.

An agent can launch a game, tool, or website and appear competent while avoiding the interactions that define the task. The difference becomes obvious in a game environment: booting the program is easy; playing requires a continuing loop of observation, action, consequence, and adaptation.

The same pattern appears in practical software. A workflow is not complete because the right screen appeared. The agent must manipulate state, notice feedback, recover from an unexpected response, and verify the final outcome. Environments that reward only milestones at the beginning of this sequence overestimate capability.

Ground variation in what people actually do.

Purely synthetic variation is cheap, but it tends to reproduce the assumptions of the generator. Real usage traces reveal categories of behavior that designers did not know to request. Support tickets, failure reports, screen recordings, and observed workflows can supply the structural patterns without requiring every task to replay private data.

The goal is not to copy individual users. It is to extract reusable forms of irregularity: interrupted sequences, ambiguous goals, unusual input sizes, conflicting constraints, and mistaken actions followed by correction. Those patterns can seed controlled task families that are both reproducible and meaningfully diverse.

Cost still matters. High-fidelity simulation is expensive, so realism should be allocated where behavior changes the result. A detailed visual world is wasteful if a simple state machine preserves the same challenge. A minimal interface is misleading if spatial interaction is the capability being tested.

Design for entropy without giving up control.

1. 01
   Collect behavioral primitives.

   Build a library of real interruptions, errors, reversals, and edge cases.

2. 02
   Compose multiple paths.

   Allow different action sequences to reach the same legitimate goal.

3. 03
   Vary state, not only wording.

   Change timing, history, constraints, and tool responses.

4. 04
   Score recovery.

   Reward noticing, correcting, and communicating about failure.

Diversity is valuable when it exercises a capability, not when it merely decorates the prompt. The environment should surprise the agent in ways that resemble the world outside the test.